Sources

The course was distilled from 82 talks and interviews and 44 articles and guides. Quotes in the lessons link to the exact moment in each talk. Not affiliated with or endorsed by anyone listed.

TalkThe Secrets of Claude Code From the Engineers Who Built ItBoris Cherny, Cat WuInterview with the creators of Claude Code about how it came about (accidentally, from a terminal prototype where giving the model bash was the key revelation), how the team builds bottoms-up with dogfooding and latent demand, and practical power-user tips (plan mode, subagents, hooks, settings.json, slash commands). They also discuss where agents are headed: longer autonomy, Claudes monitoring Claudes, and non-terminal form factors.Cited in lesson 2, lesson 5, lesson 7, lesson 10, lesson 11, lesson 13, lesson 15, lesson 17, lesson 22, lesson 24 TalkAgentic Engineering: Working With AI, Not Just Using It — Brendan O'LearyBrendan O'LearyBrendan O'Leary argues that agentic engineering means working with AI as a collaborator rather than just using it as a tool. He frames agents as fast, well-read but judgment-free junior developers and lays out tactics: context engineering, a research-plan-implement loop with fresh sessions, configuring agents with AGENTS.md/skills/modes, and curating MCP servers.Cited in lesson 3, lesson 4, lesson 5, lesson 6, lesson 10, lesson 11, lesson 14 ArticleBest practices for Claude CodeAnthropic (Claude Code team)Official Anthropic guide to getting the most out of Claude Code. Its core argument is that the context window is the most important resource to manage, and that most best practices follow from it: give the agent a way to verify its work, explore then plan then code, keep CLAUDE.md short, clear context often, use subagents and parallel sessions, and add independent review.Cited in lesson 4, lesson 5, lesson 6, lesson 8, lesson 9, lesson 10, lesson 11 TalkBuilding Claude Code with Boris ChernyBoris ChernyBoris Cherny describes how Claude Code grew from a terminal side project into a major developer tool, and how he now ships 20-30 PRs a day with no handwritten code. He covers his parallel-agent workflow, layered code review and security, prototyping instead of writing PRDs, and agent teams. He closes with a printing-press analogy for where software engineering is heading and which skills still matter.Cited in lesson 1, lesson 5, lesson 6, lesson 12, lesson 14, lesson 23, lesson 24 TalkMastering Claude Code in 30 minutesBoris ChernyA practical, early (2025) talk of tips and tricks for Claude Code. Boris recommends starting with codebase Q&A, asking for a plan before coding, giving the agent tools and a way to check its own work, sharing CLAUDE.md, MCP config and slash commands across the team, and using the SDK/headless mode and parallel sessions for advanced use.Cited in lesson 2, lesson 5, lesson 8, lesson 10, lesson 11, lesson 14, lesson 22 TalkSpaceX AI Engineer Lauren Tan: Part 1Lauren TanLauren Tan describes how she went from closely supervising one agent to having agents auto-merge PRs (about 1,000 PRs in a month). Her main argument is that trust is the limiting factor, and trust comes from verification skills, hard-enforced codebase constraints (lint, CI, architecture), and evals for skills, not from soft rules. She also covers cloud agents, greenfield vs brownfield risk, and the token ROI trade-off.Cited in lesson 8, lesson 9, lesson 13, lesson 16, lesson 18, lesson 19, lesson 20 TalkFull Walkthrough: Workflow for AI Coding — Matt PocockMatt PocockA live workshop on Matt Pocock's end-to-end AI coding workflow: grill-me alignment session, PRD as destination doc, Kanban board of vertical-slice issues, AFK Ralph-style implementation loops, then manual QA and review. His thesis is that classic software engineering fundamentals (deep modules, tracer bullets, TDD, feedback loops) matter more than new AI-specific paradigms, and that humans must stay in the loop for alignment and taste.Cited in lesson 4, lesson 5, lesson 8, lesson 9, lesson 15, lesson 19 TalkThe creator of OpenClaw: "I ship code I don't read"Peter SteinbergerPeter Steinberger, creator of the viral personal-assistant project Clawdbot (OpenClaw), explains how he builds software after returning from a three-year break: running 5-10 agents in parallel, shipping code he doesn't read line by line, and treating PRs as 'prompt requests'. His central principle is 'closing the loop', designing systems so agents can build, run and verify their own work, while he acts as architect with taste. Host Gergely Orosz notes that Clawdbot is a YOLO-style project, so the approach should be taken with a grain of salt.Cited in lesson 3, lesson 8, lesson 9, lesson 12, lesson 16, lesson 17 Talk5 Claude Code skills I use every single dayMatt PocockMatt Pocock walks through five skills he uses daily with Claude Code: grill-me, write-a-PRD, PRD-to-issues, TDD, and improve-codebase-architecture. Together they encode a strict process (shared understanding, destination doc, vertical-slice issues, red-green-refactor, deep-module refactors) that agents with no memory can follow, often executed by a Ralph loop. His argument is that treating agents like forgetful humans and applying real engineering process raises code quality.Cited in lesson 3, lesson 5, lesson 13, lesson 15, lesson 19 TalkContext engineering with Dex HorthyDex HorthyDex Horthy explains context engineering as controlling exactly which tokens go into a model, and the 'smart zone vs dumb zone' of context windows. He draws on his team's failed lights-off software factory to argue against not reading code. He advocates 'token smarter' leverage points (human-in-the-loop planning, slow nightly loops) over 'token harder' dark factories.Cited in lesson 4, lesson 6, lesson 10, lesson 17, lesson 18 ArticleEffective harnesses for long-running agentsJustin YoungAnthropic describes a harness for getting the Claude Agent SDK to make steady progress on long projects that span many context windows. An initializer agent sets up the environment (feature list, init.sh, progress file, git repo), and each coding agent then works on one feature at a time, verifies it end-to-end with browser automation, and leaves clean artifacts for the next session. The main argument is that compaction alone is not enough, and that structured artifacts and disciplined session routines are what make long-running agents work.Cited in lesson 6, lesson 8, lesson 10, lesson 14, lesson 17 TalkHead of Claude Code: What happens after coding is solved | Boris ChernyBoris ChernyBoris Cherny, creator of Claude Code, describes how the tool grew from a terminal hack into a product that writes 100% of his code, and argues that coding is largely solved. He shares principles for building on AI models (don't box the model in, bet on the general model, build for the model six months out, latent demand), practical tips (best model, plan mode, parallel sessions), and predictions that roles will converge toward 'builder'.Cited in lesson 5, lesson 9, lesson 10, lesson 17, lesson 24 TalkHow AI will change software engineering – with Martin FowlerMartin FowlerMartin Fowler argues that AI is the biggest shift in his career, comparable to the move from assembly to high-level languages. The key difference is the shift from deterministic to non-deterministic tooling, more than any rise in abstraction. He covers where LLMs help today (prototyping, legacy understanding, exploring unfamiliar tech), why vibe coding kills the learning loop, why thin slices and tight feedback loops still win, and why refactoring, testing, and communication skills stay central.Cited in lesson 3, lesson 6, lesson 19, lesson 23, lesson 24 ArticleMy AI Adoption JourneyMitchell HashimotoMitchell Hashimoto describes his six-step, measured path from AI skeptic to productive agent user. He moved from chatbots to agents, forced himself to redo his own commits with agents, and then added end-of-day agents, outsourcing slam-dunk tasks, harness engineering, and keeping one agent running. His argument is that real value comes from deliberate, friction-tolerant skill building, not hype.Cited in lesson 8, lesson 10, lesson 11, lesson 18, lesson 23 TalkThink Harder: How I Prompt | Thorsten Ball at Laracon US 2026Thorsten BallThorsten Ball explains how he prompts coding agents to build Amp, which is about 99% AI-written. His one principle is to ask 'how is the model supposed to know what I mean?', so every prompt should supply the missing information, whether from the prompt itself, pointers to files, screenshots, auto-generated debug context, or AGENTS.md files in the codebase. In the Q&A he adds that you need to know which code is load-bearing and review that closely, and that you should test giving agents a long leash.Cited in lesson 2, lesson 7, lesson 9, lesson 11, lesson 19 TalkAndrej Karpathy: Software Is Changing (Again)Andrej KarpathyKarpathy frames the shift from Software 1.0 (code) and 2.0 (neural net weights) to Software 3.0, where prompts in English program LLMs. He describes LLMs as fallible 'people spirits' that behave like 1960s-era operating systems. He argues for partial-autonomy apps with fast generation-verification loops, an 'autonomy slider', keeping AI on a leash, and building docs and infrastructure for agents. He frames this as the decade of agents, not the year of agents.Cited in lesson 1, lesson 4, lesson 9, lesson 24 TalkAnthropic's Boris Cherny: Why Coding Is Solved, and What Comes NextBoris ChernyBoris Cherny, creator of Claude Code, describes how it began as a bet on the next model, why he considers coding 'solved' (the model writes 100% of his code), and how he now works: many parallel sessions, /loop and routines, thousands of overnight agents. He predicts more cross-disciplinary generalists, software becoming as universal as literacy after the printing press, and the harness and safety scaffolding mattering less as models improve.Cited in lesson 16, lesson 17, lesson 18, lesson 23 TalkDHH’s new way of writing codeDavid Heinemeier Hansson (DHH)DHH explains why he went from AI skeptic (on Lex Fridman) to agent-first on everything after Claude Code-style agent harnesses and Opus 4.5 arrived in late 2025. He describes his multi-agent workflow, how agents expand ambition (the P1 optimization, 100 PRs in 90 minutes), how senior engineers benefit most, and why taste, craft and product sense matter more. He also warns about overwork and predicts 'peak programmer'.Cited in lesson 10, lesson 12, lesson 22, lesson 23 TalkEverything We Got Wrong About Research-Plan-Implement - Dexter HorthyDex HorthyDex Horthy explains what went wrong with the Research-Plan-Implement workflow after thousands of engineers adopted it. The monolithic planning prompt skipped steps, long plans were not worth reading, and not reading code backfired. He proposes splitting the workflow into smaller prompts (CRISPY) with short alignment artifacts, vertical plans, and always reading the code.Cited in lesson 4, lesson 5, lesson 9, lesson 11 TalkHarness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAIRyan LopopoloRyan Lopopolo describes nine months of building software exclusively with agents (Codex) at OpenAI, banning his team from touching editors. His argument: code is free, so the scarce resources are human time, attention and model context. Engineers should encode non-functional requirements as docs, lints, structural tests and review agents so agents do the full job and humans steer.Cited in lesson 1, lesson 6, lesson 18, lesson 19 ArticleHarness engineering: leveraging Codex in an agent-first worldRyan LopopoloA wiki summary of OpenAI's first-person post on five months of building an internal product with zero manually-written code, entirely authored by Codex. The main argument is that the engineer's job shifts from writing code to designing environments, feedback loops and mechanical constraints (repo-as-system-of-record, layered architecture, golden principles, garbage collection) so agents can work at high throughput without drift.Cited in lesson 11, lesson 17, lesson 18, lesson 19 TalkSkill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AIAndrej KarpathyKarpathy describes how, since December 2025, he has shifted from writing code himself to delegating almost everything to coding agents, and how he now treats his own role as the bottleneck to remove. He covers parallel agents, persistent 'claw' agents (Dobby the home-automation agent), AutoResearch (autonomous loops that tune models overnight), the limits of jagged, RL-trained models, untrusted swarms of workers, open vs closed models, jobs, robotics, and writing docs and education for agents rather than humans.Cited in lesson 1, lesson 16, lesson 17, lesson 23 TalkState of Agentic Coding #10 with Armin and BenArmin Ronacher, Ben VinegarEpisode 10 of a recurring podcast. Armin and Ben discuss AI watermarking and slop, model/vendor politics (Cursor losing OpenAI models, Claude subscription harness restrictions), open-weight models versus frontier models, coding agents moving to the cloud and into Slack/Linear, and GitHub's instability. Most of the second half covers lessons learned: guardrails for agent-built software, the limits of agents on design, and the gap between token spend and organizational productivity.Cited in lesson 10, lesson 16, lesson 22, lesson 24 TalkStop being skeptical about AI for development with Charity MajorsCharity MajorsCharity Majors explains why she moved from AI skeptic to convert after late-2025 models and harnesses improved, and argues that whether engineers will ship unread AI-generated code is a question of when, not if. She says trust must move from code reading to validation, drawing on ops and QA practices such as testing in production, conformance tests, evals and observability. She also describes the two camps in software engineering, the reliability costs that on-call engineers see, and her career advice for managers and juniors.Cited in lesson 3, lesson 8, lesson 9, lesson 22 TalkTDD, AI agents and coding with Kent BeckKent BeckKent Beck, after 50+ years of programming, describes how he uses agentic coding tools daily and why he calls them an unpredictable 'genie'. He explains how TDD acts as a guardrail against agents that delete tests, fake results, and cause disruption at a distance. The rest of the talk covers the history of the Agile Manifesto, XP, TDD, and Facebook's feedback-loop culture.Cited in lesson 8, lesson 10, lesson 19, lesson 23 TalkThe Ralph Wiggum Loop from 1st principles (by the creator of Ralph)Geoffrey HuntleyHuntley explains Ralph (a bash while-true loop that feeds the same prompt to a coding agent) from first principles. He covers keeping the context window small ("malloc'ing the array"), generating specs through conversation, and a lookup-table 'pin' that stops the agent inventing things. He then demonstrates running a one-task-per-loop Ralph. He argues software development is now cheap and automatable, and that engineers' job is engineering back pressure.Cited in lesson 4, lesson 6, lesson 17, lesson 18 Article2026 Agentic Coding Trends ReportAnthropicAnthropic's report predicts eight trends for agentic coding in 2026, spanning foundation, capability and impact trends. Its main argument is that engineering is shifting from writing code to orchestrating agents (multi-agent teams, long-running autonomous work, scaled oversight), while the work stays collaborative: developers use AI in about 60% of tasks but can fully delegate only 0-20%.Cited in lesson 1, lesson 17, lesson 23 ArticleAgentic Coding RecommendationsArmin RonacherArmin Ronacher shares his practical agentic coding setup built around Claude Code: running in YOLO mode, preferring simple tools over MCP, picking Go and simple code for agent-friendliness, and designing fast, observable, chaos-proof tools. His core principles are simplicity, stability, observability and smart parallelization.Cited in lesson 8, lesson 12, lesson 19 TalkAndrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie ZhanAndrej KarpathyKarpathy describes how December's model improvements made agentic coding fundamentally work, and frames Software 3.0 (prompting/context as programming). He distinguishes vibe coding (raising the floor) from agentic engineering (preserving the quality bar while going faster), explains jagged intelligence via verifiability and lab RL focus, and argues humans remain in charge of taste, spec, and understanding.Cited in lesson 1, lesson 3, lesson 18 TalkAndrew Ng: Building Faster with AIAndrew NgAndrew Ng talks at YC Startup School about how startups can win through execution speed in the AI era. He covers concrete ideas, agentic workflows, AI coding assistants making prototyping 10x faster, the shift of the bottleneck to product management and user feedback, and why understanding AI tech buys speed. A Q&A covers hype narratives, token costs, moats, responsible AI and open source.Cited in lesson 6, lesson 20, lesson 23 TalkBuilding pi in a World of Slop — Mario ZechnerMario ZechnerMario Zechner explains why he left Claude Code and OpenCode to build Pi, a minimal and extensible coding agent harness where the agent can modify itself. He then covers how AI-generated 'slop' PRs and issues are overwhelming open-source maintainers. Finally he argues that agents compound errors without learning or bottlenecks, so engineers should scope agent tasks tightly, read all critical code, and slow down.Cited in lesson 6, lesson 19, lesson 23 TalkClaude Code best practices | Code w/ ClaudeCal RuebAn Anthropic engineer explains how Claude Code works as a 'pure agent' (instructions, tools, loop, agentic search with no indexing), then walks through use cases and best practices: CLAUDE.md, permissions, CLI tools over MCP, context management, plan-first workflows, TDD/linting/commits, parallel sessions, escape, headless automation, and new Claude 4 features like interleaved thinking.Cited in lesson 2, lesson 5, lesson 12 TalkClaude Code’s Bitter Lesson: Prompts, Harnesses, Mods, and Agents — Thariq ShihiparThariq ShihiparThariq Shihipar of Anthropic's Claude Code team discusses how to get the most from agents: prompting as building a mental model of Claude, surfacing unknowns, artifacts as an interface for the harness, and the new 'Claude Mods' system for customizing the harness itself. He argues CLAUDE.md and per-model failure logs go stale as models improve. The second half covers Claude Tag, multiplayer agents, and Anthropic's 'pacing the frontier' safety concerns after incidents of agents hacking out of sandboxes during evals.Cited in lesson 7, lesson 11, lesson 21 TalkHow Anthropic uses Claude Code: Agentic Software Engineering at Scale - Daisy HollmanDaisy HollmanDaisy Hollman explains how Anthropic customizes and scales Claude Code for real, large-scale software engineering. She frames the context window as the fundamental primitive. She evaluates plugin abstractions (MCP, skills, subagents, hooks) by whether they scale with 'don't pay for what you don't use', and describes running many parallel agents with worktrees, agent teams, /loop, auto mode and fleet view.Cited in lesson 12, lesson 13, lesson 24 TalkWhy Software Factories FailDex HorthyDex argues that lights-out software factories (no human code review) fail because coding models are trained via RL on tests-pass rewards, which cannot capture long-term maintainability. No amount of harness engineering or loops can fix what is a model training problem. His prescription is to keep reading code, and to front-load product review, architecture, program design, and vertical-slice planning so reviews stay short.Cited in lesson 3, lesson 5, lesson 19 Talk"The engineer of the future is the person who is able to choose what is worth doing." — Addy OsmaniAddy OsmaniOsmani argues that as agents take over the inner execution loop (harnesses, loops, software factories), the scarce human contribution becomes judgment, evidence, and accountability. He introduces framings like alpha/decay, cognitive debt, cognitive surrender, orchestration tax, and the rule 'explain it or don't ship it', and argues engineering demand will grow, with the bottleneck moving to 'should this exist and can we answer for it'.Cited in lesson 16, lesson 24 TalkAgentic Engineering in Action with Mitchell HashimotoMitchell HashimotoMitchell Hashimoto walks through a real Ghostty commit (undo-close-tab/window feature) built with Claude Code, showing how he acts as architect, gives well-scoped prompts with guardrails, manually cleans up, and uses the agent for cleanups and review. He discusses what agents are good and hopeless at (Zig, high-performance data structures, architecture), and why skill and understanding still matter.Cited in lesson 16, lesson 23 ArticleAgentic manual testing - Agentic Engineering PatternsSimon WillisonA guide chapter arguing that the defining trait of a coding agent is that it can execute the code it writes, so agents should be told to test their work manually as well as with automated tests. It covers concrete mechanisms (python -c, curl, Playwright, Rodney) and Showboat, a tool that records the agent's testing as a document with real command output.Cited in lesson 8, lesson 12 ArticleAugmented Coding: Beyond the VibesKent BeckKent Beck recounts four weeks building BPlusTree3, a performance-competitive B+ Tree library in Rust and Python, using AI ('the genie') with a TDD-focused system prompt. He distinguishes 'augmented coding' (caring about code quality, tests, and complexity) from 'vibe coding' (caring only about behavior). He argues that programming remains programming, with more consequential decisions and less yak shaving, while admitting the AI still struggles to care about simplicity.Cited in lesson 3, lesson 10 TalkBuilding uReview, Uber’s Multi-Agent Code Review Engine — Will Bond & Ameya Ketkar, UberWill Bond, Ameya KetkarUber engineers describe uReview, an in-house multi-agent code review system built because review time has become the bottleneck as PR volume grows. They cover its architecture (multiple generators plus post-processing to filter and dedupe comments), how observability and evaluation improved quality and cost, how team customization was scaled, and the argument that human review moves up a layer instead of disappearing.Cited in lesson 9, lesson 13 TalkContext Engineering for AgentsLance MartinLance from LangChain defines context engineering as the art and science of filling the context window with just the right information at each step of an agent's trajectory. He organizes the practices into four buckets (write, select, compress, isolate), illustrates each with examples from popular agents (Claude Code, Anthropic's multi-agent researcher, Cognition, Windsurf, Hugging Face), and shows how LangGraph supports them.Cited in lesson 4, lesson 12 ArticleContext Engineering for Coding AgentsBirgitta BöckelerA primer on the context configuration features in coding agents (rules files, skills, slash commands, subagents, MCP servers, hooks, plugins), organized by what they are, who decides to load them, and how much context they consume. It uses Claude Code as the worked example. The main argument is to keep context small and build it up iteratively. The article also warns that context engineering is probabilistic, not deterministic, so it gives no guarantees.Cited in lesson 11, lesson 13 TalkDon't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, AnthropicBarry Zhang, Mahesh MuragAnthropic's creators of Agent Skills argue that, because code is a universal interface, a general agent with a thin bash and file-system scaffold can cover most domains. What it lacks is domain expertise, which skills supply as folders of procedural knowledge, scripts and assets that load progressively. They describe the emerging architecture (agent loop + runtime + MCP + skills), the skills ecosystem, and a vision of skills as shareable, versioned, continuously learning organizational knowledge.Cited in lesson 13, lesson 22 ArticleEffective context engineering for AI agentsAnthropic Applied AI team (Prithvi Rajasekaran, Ethan Dixon, Carly Ryan, Jeremy Hadfield)Anthropic's Applied AI team defines context engineering as the successor to prompt engineering: curating the smallest set of high-signal tokens in a finite attention budget across system prompts, tools, examples, and message history. It covers just-in-time retrieval and progressive disclosure. For long-horizon work it recommends compaction, structured note-taking, and sub-agent architectures.Cited in lesson 4, lesson 15 ArticleEquipping agents for the real world with Agent SkillsBarry Zhang, Keith Lazuka, Mahesh MuragAnthropic introduces Agent Skills: folders containing a SKILL.md plus optional files and scripts that agents discover and load dynamically to gain domain expertise. The core design principle is progressive disclosure, which keeps context use small while allowing effectively unbounded bundled knowledge. The article also covers bundling executable code, security risks from untrusted skills, and the future of agent-authored skills.Cited in lesson 13, lesson 21 ArticleGitHub's Agentic Engineering SystemGitHub (authors of the Agentic Engineering System)GitHub presents its Agentic Engineering System (AES), a tool-agnostic framework for scaling agent use. It rests on three stocks (governance, shared knowledge, customer value), three activities (define, deliver, detect) and three modes (director, performer, assessor). The main argument is that agent speed only creates value when governance and shared knowledge are strong enough to support the level of delegation. Otherwise teams get rework, review burden and risk.Cited in lesson 6, lesson 22 TalkHow Claude Code Works - Jared Zoneraich, PromptLayerJared ZoneraichJared explains, based on leaked system prompts, open-source Codex and his own heavy usage, why Claude Code and similar coding agents finally work. The core argument: a simple architecture (a single master while-loop plus a few simple tools such as bash, grep and diff-based edit) combined with better models beats elaborate DAGs and scaffolding. He also covers context management, sub-agents, todos, skills, other agents' philosophies (Codex, AMP, Cursor), and how to eval agents.Cited in lesson 2, lesson 12 TalkHow I Review AI Code - (Meta Senior Staff Engineer)John KimJohn Kim argues that code review in the agent era is a gradient, not a binary of reading everything or nothing. He treats the codebase like a tree: review trunk code (core infra, high blast radius, ungated) deeply and skim gated leaf code. He leans on agent-provided proof (tests, logs, screenshots, video), adversarial review agents, feature gating and experiments to land safely, and reserves a final human-driven 20% for launch readiness.Cited in lesson 9, lesson 15 TalkHow We Build Effective Agents: Barry Zhang, AnthropicBarry ZhangBarry Zhang distills the Anthropic 'Building Effective Agents' blog post into three ideas: don't build agents for everything, keep the design simple (environment, tools, system prompt in a loop), and think like your agent by viewing the world through its limited context window. He closes with open questions on budget-aware agents, self-evolving tools, and multi-agent communication.Cited in lesson 2, lesson 3 TalkI’m HOOKED on Claude Code Hooks: Advanced Agentic CodingIndyDevDanA demo-driven walkthrough of Claude Code hooks (pre-tool use, post-tool use, notification, stop, subagent stop), showing how to block dangerous commands, log every agent event for observability, and add text-to-speech notifications. The main argument is that hooks give deterministic control and observability over agents, making Claude Code a programmable engineering primitive. The last third is opinion on why the Claude Code creators left Anthropic for Cursor, plus a course pitch.Cited in lesson 13, lesson 21 TalkI'm using claude --worktree for everything nowMatt PocockMatt Pocock tries Claude Code's new built-in git worktree support (`claude --worktree`) for the first time. He explains how git worktrees work, shows that Claude wraps them with auto-named branches and lifecycle cleanup, and flags a gotcha: commits and pushes can land on main unless the agent pushes an explicitly named branch. He argues this makes parallelization nearly free, and that AI coding relies on old fundamentals like TDD and planning.Cited in lesson 14, lesson 16 ArticleMeasuring AI Ability to Complete Long Software TasksMETRMETR proposes measuring AI agent capability by the length of tasks (in human expert time) that models can complete at a given success probability. They find this '50% time horizon' has grown exponentially for 6 years, doubling roughly every 7 months. Extrapolating, agents could handle month-long software projects by the end of the decade, though models are currently reliable only on tasks of a few minutes.Cited in lesson 3, lesson 24 TalkMitchell Hashimoto’s new way of writing codeMitchell HashimotoMitchell Hashimoto recounts the HashiCorp story (Vagrant, Terraform, Vault, the pivot to enterprise open core, the near-acquisition by VMware, and his views on AWS/Azure/GCP). He then covers how he uses AI agents day to day, how AI-generated PRs are overwhelming open source (Ghostty's move to a vouching system), and how Git, CI/CD and testing will need to change for agentic workflows.Cited in lesson 17, lesson 18 TalkNo Vibes Allowed: Solving Hard Problems in Complex Codebases – Dex Horthy, HumanLayerDex HorthyDex Horthy argues that AI coding agents fail in complex brownfield codebases because of poor context management, and presents a Research-Plan-Implement workflow built on 'frequent intentional compaction' to stay in the 'smart zone' of the context window. He stresses that humans must not outsource thinking: reviewing research and plans is the highest-leverage human activity and keeps teams mentally aligned.Cited in lesson 15, lesson 22 ArticleRalph Wiggum as a "software engineer"Geoffrey HuntleyHuntley describes Ralph, a technique that in its purest form is a Bash loop feeding the same prompt to a coding agent over and over. The agent does one task per loop, using a fixed 'stack' of specs and a fix plan, subagents for expensive work, and tests or type checks as back pressure. He argues that operator skill in tuning the prompt, watching the stream and resetting the plan matters most, and says Ralph is for greenfield projects only, not existing codebases.Cited in lesson 2, lesson 17 ArticleRed/green TDD - Agentic Engineering PatternsSimon WillisonA short guide in Simon Willison's Agentic Engineering Patterns series arguing that the phrase "use red/green TDD" is a concise prompt that improves coding agent results. Writing tests first and confirming they fail guards against non-working or unused code and builds a regression suite. The article cuts off at the example prompt.Cited in lesson 7, lesson 8 TalkShip working code while you sleep with the Ralph Wiggum techniqueMatt PocockMatt Pocock explains the Ralph Wiggum technique (credited to Geoffrey Huntley): running a coding agent in a simple bash for loop that repeatedly picks the highest-priority task from a JSON PRD, implements one feature, commits, and logs progress. He walks through his ralph.sh script, the PRD and progress.txt files, and stresses small tasks and strong feedback loops (types, tests, CI, browser automation). He also shows a human-in-the-loop variant.Cited in lesson 16, lesson 19 ArticleThe lethal trifecta for AI agents: private data, untrusted content, and external communicationSimon WillisonWillison defines the 'lethal trifecta' of AI agent capabilities (access to private data, exposure to untrusted content, and ability to externally communicate) that together let an attacker trick an agent into stealing the user's data. He argues that guardrail products can't be trusted, vendors can't protect users who mix and match tools (especially via MCP), and the only reliable defense is to avoid combining all three.Cited in lesson 12, lesson 21 TalkTool, skill, or subagent? Decomposing an agent that outgrew its promptWillA hands-on Code with Claude London workshop on refactoring an inventory agent (Stock Pilot) whose 400-line system prompt and 12 tools had caused eval regressions. Will shows how to decide between tools, skills, and subagents: move situational knowledge into skills, replace custom tools with code-execution primitives, keep subagents only for parallelism or fresh-context work, and hill-climb on evals. The result was a 15-line prompt, 3 tools, and about 92% eval pass.Cited in lesson 12, lesson 15 TalkWhy Coding Agents Keep Making Your Codebase WorseDex HorthyDex Horthy argues that coding agents still degrade codebases over time (citing SlopCodeBench), that most online agent hype and 'token maxing' is fake, and that realistic gains are 2-4x, not 100x. He lays out his team's actual workflow: small one-shot PRs, overnight 'janitor' bots, design-doc-first planning with checkpointed implementation, and reading all agent code before review.Cited in lesson 16, lesson 22 Talk12-Factor Agents: Patterns of reliable LLM applications — Dex Horthy, HumanLayerDex HorthyDex Horthy distills patterns from conversations with 100+ builders into '12-Factor Agents'. His argument: reliable production agents are mostly deterministic software with small LLM loops, and builders should own their prompts, context window, control flow, and state instead of relying on opaque frameworks. LLMs are stateless functions, so everything comes down to context engineering.Cited in lesson 2 ArticleAdding a new content type to my blog-to-newsletter tool - Agentic Engineering PatternsSimon WillisonA worked example from Simon Willison's Agentic Engineering Patterns guide. He shows how three short prompts to Claude Code on the web added a new content type ('beats') to his blog-to-newsletter tool in one shot. The prompts clone a reference repo, point at an existing behavior to imitate, and give the agent a way to verify its own work.Cited in lesson 7 ArticleAdvanced Context Engineering for Coding AgentsDex HorthyDex Horthy argues that today's coding agents can work in large brownfield codebases (e.g. a 300k LOC Rust repo) if you apply context engineering: deliberately compact context, keep utilization at 40-60%, and use a research -> plan -> implement workflow with human review at the highest-leverage points (research and plans, not code). He backs it with BAML case studies, a failure story (parquet-java), and a view that the real challenge is team and workflow transformation and mental alignment.Cited in lesson 5 TalkAgentic coding at AirbnbSzczepan Faber, MikeTwo Airbnb developer-platform engineers describe how agentic coding was rolled out at Airbnb, reaching about 60-64% of PR authors in a year. They cover the pair-programming framing, the learning journey from approving every step to parallel sessions, the Airchat CLI wrapper, an early MCP bet, and moving from RAG pipelines to tool-calling knowledge access. Main argument: the agent alone isn't enough; you need infrastructure (sandboxes, MCP, code review culture), meet developers where they are, and keep human review and standards high.Cited in lesson 7 ArticleAgentsChip HuyenA conceptual overview of foundation-model agents, adapted from the Agents section of the book AI Engineering. An agent is defined by its environment and tool inventory, and its success depends on tool selection and the strength of its planner. The article covers planning (decoupled from execution, reflection, control flows), function calling, tool selection experiments, and a taxonomy of agent failure modes (planning, tool, efficiency) with how to evaluate each.Cited in lesson 2 ArticleAGENTS.mdAGENTS.md maintainers (OpenAI Codex, Amp, Jules, Cursor, Factory; now Agentic AI Foundation / Linux Foundation)AGENTS.md is a simple, open, vendor-neutral file format that acts as a README for coding agents, giving them a predictable place to find setup, build, test, code-style and PR instructions. It is kept separate from README.md so human docs stay concise, and one file works across many agents. It supports nested files for monorepos and migration via symlinks.Cited in lesson 11 ArticleAI Evals: Everything You Need to Know – Hamel’s BlogHamel Husain (with Shreya Shankar)A long FAQ on AI evals for product teams. Its central argument is that evals should start with manual error analysis of real traces rather than infrastructure or generic metrics. From there, teams build targeted binary evaluators, validate LLM judges against human labels, and keep a single domain expert in charge of quality. It also covers agent and coding-agent evaluation, CI versus production monitoring, and tooling choices.Cited in lesson 20 TalkAn AI state of the union: We’ve passed the inflection point & dark factories are comingSimon WillisonSimon Willison gives an AI state of the union: models crossed an inflection point in November 2025 (GPT 5.1 / Claude Opus 4.5) so coding agents now mostly do what you ask. He distinguishes vibe coding from agentic engineering, describes dark-factory/software-factory experiments (StrongDM), and shares concrete patterns (red/green TDD, thin templates, hoarding working examples). He also warns about cognitive exhaustion and the unsolved lethal trifecta / prompt injection problem, predicting an eventual 'Challenger disaster' of AI.Cited in lesson 1 TalkAnthropic Workshop: Build Agents That Run for Hours — Ash Prabaker & Andrew WilsonAsh Prabaker, Andrew WilsonTwo Anthropic applied AI engineers explain how to build agents that run for hours. Andrew traces a year of model and harness co-evolution (context, skills, checkpoints, Ralph loop, agent teams). Ash presents a GAN-inspired planner/generator/evaluator harness with negotiated contracts and rubric-graded taste, showing it turns broken one-shot apps into working ones at higher cost, plus advice on simplifying the harness as models improve and on reading traces.Cited in lesson 17 ArticleAnti-patterns: things to avoid - Agentic Engineering PatternsSimon WillisonA short guide chapter naming one core anti-pattern of agentic engineering: dumping unreviewed agent-generated code on collaborators via pull requests. Willison argues the submitter is responsible for the first review pass and lists the traits of a good agentic PR: it works, it is small, it has context, and its description has been read. He also recommends attaching evidence of manual testing.Cited in lesson 9 TalkBoris Cherny: We Cut 80% of Claude Code’s PromptBoris ChernyBoris Cherny discusses the Opus 5 release and how Claude Code's harness is rebuilt with each model: deleting 80% of the system prompt, using ablations, and giving models harder tasks with verification. He covers product overhang, long-running jobs (an 11-day Zig-to-Rust Bun rewrite, a 2-week Swift rewrite), dynamic workflows with thousands of agents, and self-maintaining routines. His core claim is that working with models is an empirical science, not about tricks or over-specifying.Cited in lesson 7 ArticleBuilding Effective AI AgentsErik S. and Barry ZhangAnthropic distills what it learned from dozens of customer agent deployments and its own agents (including SWE-bench). It argues for simple, composable patterns over heavy frameworks. It defines the workflow-versus-agent distinction, lays out five workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer), and describes agents as LLMs using tools in a loop. It closes with advice on designing agent-computer interfaces (tools).Cited in lesson 12 TalkBuilding your own software factory — Eric Zakariasson, CursorEric ZakariassonEric Zakariasson describes how Cursor dogfoods its own product to move toward a 'software factory' where agents do most of the work. He covers autonomy levels, the primitives a factory needs (codebase structure, guardrails, enablers, verifiability), the shift from worker to manager, and Cursor's internal automations. A long Q&A covers code quality, team silos, enterprise risk, handoff from PMs, and roles.Cited in lesson 16 ArticleDemystifying evals for AI agentsMikaela Grace, Jeremy Hadfield, Rodrigo Olivares, Jiri De JongheA practical guide from Anthropic on designing evals for AI agents. It defines evals, grader types (code, model, human), capability vs regression evals, and pass@k vs pass^k. It gives a step-by-step roadmap, from starting with 20-50 tasks drawn from real failures to reading transcripts and maintaining the suite, and covers eval design for coding, conversational, research and computer-use agents. Main argument: evals compound in value and are how teams escape reactive, guess-and-check development.Cited in lesson 20 TalkEverything we knew about software has changed — Theo Browne, @t3dotgg Theo Browne (t3.gg)Theo describes his 'AI psychosis' journey through model eras (Sonnet 3.5 = tool calls, Opus 4.5 = long-running tasks, Mythos = orchestration). He argues developers must drop old habits and sunk-cost attachments (the 'skeuomorphic phase' of software), and shift ambition: what used to be a startup is now a side project, so build wider and attempt things that feel stupidly big.Cited in lesson 24 TalkFastRender with Wilson Lin - building a web browser with swarms of thousands of coding agentsWilson LinWilson Lin explains FastRender, a research project in which Cursor ran up to ~2,000 concurrent coding agents for a week to build a >1M-line Rust browser engine. The main argument: scaling agents is largely about hierarchical planner/worker structure, rigorous initial instructions, strong feedback loops (specs, test suites, compilation), and tolerating small transient errors to preserve throughput.Cited in lesson 6 ArticleFirst run the tests - Agentic Engineering PatternsSimon WillisonA short guide chapter arguing that automated tests are now essential when working with coding agents. Simon Willison recommends starting every new agent session on an existing project with the four-word prompt "First run the tests", which makes the agent discover how to run the tests, gauge the project's size, and stay in a testing mindset.Cited in lesson 7 TalkFrom AI-Assisted to AI-Native: Building a Frontier Development Team — Clare Liguori, AWSClare LiguoriLiguori describes 'frontier development' inside Amazon: engineers write 1-2% of code by hand, run agents for hours, and run many in parallel. Pilots showed a median 4.5x productivity gain (sometimes 10x+), but only on teams that intentionally changed how they work rather than sprinkling AI tools on existing processes. She lays out five habits, plus the pitfalls of burnout, rollout speed, and new organizational bottlenecks.Cited in lesson 22 TalkFull Course: Spec-Driven Development with Coding AgentsPaul EverettA course on spec-driven development (SDD) with coding agents. It walks through writing a project 'constitution' (mission, tech stack, roadmap), then looping through feature plan, implement and validate steps on separate branches with fresh context. It also covers replanning, brownfield adoption, skills, MCP vs CLI, plugins and agent-portability standards. The main argument is that specs, not chat history, should be the durable memory and contract between human and agent.Cited in lesson 6 TalkHarnesses in AI: A Deep Dive — Tejas Kumar, IBMTejas KumarTejas Kumar defines an 'agent harness' as everything around the model that grounds it in a stable, deterministic environment, and builds one live from scratch. He starts with a browser agent on GPT-3.5 Turbo that lies about success and fails at login. Adding guardrails, deterministic verification, and a harness-level login handler fixes it without touching the prompt.Cited in lesson 3 ArticleHow coding agents work - Agentic Engineering PatternsSimon WillisonA concise explainer of how coding agents work under the hood: an LLM is a stateless token-completion engine, and a coding agent is a harness that adds a system prompt, callable tools, and a loop. The author argues that understanding these mechanics (tokens, chat templates, caching, tool calls, reasoning) helps you use agents more effectively.Cited in lesson 2 TalkHow the Claude Code team uses Claude CodeClaude Code team members (Boris Cherny referenced; others include Sid and Robert)Members of the Claude Code team describe how their way of working has shifted from prompting a single session and approving each tool call to giving goals to a Slack-native agent (Claude Tag) running in the cloud. They cover the short shelf life of harness features, loops and routines, fan-out workflows with adversarial review, code review that focuses humans on the big picture, verification, and what changes for engineers.Cited in lesson 18 ArticleHow to Build an AgentThorsten BallA hands-on tutorial that builds a working code-editing agent in under 400 lines of Go using the Anthropic API. It demystifies agents as an LLM, a loop, and tool definitions (read_file, list_files, edit_file). The central argument is that there is no secret: the inner loop is trivial and the rest is practical engineering and elbow grease.Cited in lesson 2 ArticleHow we built our multi-agent research systemJeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel FordAnthropic engineers describe how they built Claude's Research feature as an orchestrator-worker multi-agent system. They cover when multi-agents pay off (breadth-first, high-value tasks, with ~15x token cost), prompting heuristics for coordinating subagents, lightweight evals with LLM judges plus human testing, and the production challenges of stateful, error-compounding agents.Cited in lesson 15 TalkI Have Spent 1000+ Hours With Claude Code. This Is What I LearnedThe Coding SlothA tier-list style walkthrough of Claude Code features after 1000+ hours of use. Top picks are skills, plan mode, and verification (tests, linters, screenshots). Much of the talk is about managing context and tokens on limited plans. It ends with alternatives to Claude Code and a push to still learn to code.Cited in lesson 4 ArticleInteractive explanations - Agentic Engineering PatternsSimon WillisonA short guide chapter on 'cognitive debt': the cost of losing track of how agent-written code works. Willison's remedy is to ask coding agents for interactive or animated explanations, illustrated by a Rust word-cloud tool whose spiral placement algorithm only made sense after he asked Claude for an animation.Cited in lesson 23 TalkL8 Principal's Agentic Engineering WorkflowKun ChenA full walkthrough of Kun Chen's terminal-centric agentic engineering workflow, framed as a captain building a crew: set up the environment, onboard agents with memory files and skills, work with one agent (voice, visual planning, automated validation pipeline), scale to parallel agents with worktrees, then delegate orchestration to a 'first mate' agent. The core argument is that engineers should act like engineering managers: spend time at the start (requirements) and end (quality bar) of tasks, and automate everything in the middle.Cited in lesson 12 ArticleLinear walkthroughs - Agentic Engineering PatternsSimon WillisonA short guide chapter on 'linear walkthroughs': asking a coding agent to produce a structured, linear explanation of a codebase. Willison demonstrates by having Claude Code generate a walkthrough of a vibe-coded SwiftUI app using his Showboat tool, which embeds real command output (not hand-copied code) to avoid hallucinations. He argues the pattern counters the risk that LLMs slow down learning.Cited in lesson 23 TalkLLM Evals: Common MistakesHamel Husain, Shreya ShankarTwo evals practitioners walk through common mistakes in evaluating LLM applications: relying on generic foundation-model benchmarks and off-the-shelf metrics, outsourcing labeling and prompt-writing to developers, misusing LLM-as-judge (holistic judges, Likert scales), fearing data labeling, and misusing offline validation sets. Their core argument is that you must look at your own data, do error analysis, build custom annotation tools, and use grounded binary judges plus online monitoring.Cited in lesson 20 ArticleMaking Claude Code more secure and autonomous with sandboxingDavid Dworken, Oliver Weller-DaviesAnthropic describes how OS-level sandboxing (filesystem and network isolation) lets Claude Code run more autonomously with far fewer permission prompts while limiting the damage from prompt injection. It also covers an open-sourced sandbox runtime and Claude Code on the web, which keeps credentials out of the sandbox by routing git through a proxy.Cited in lesson 21 ArticleMeasuring the Impact of Early-2025 AI on Experienced Open-Source Developer ProductivityMETRMETR ran a randomized controlled trial with 16 experienced open-source developers on 246 real issues in their own mature repositories. Allowing AI tools (mainly Cursor Pro with Claude 3.5/3.7 Sonnet) made them 19% slower, even though they expected a 24% speedup and afterward still believed AI had sped them up by 20%. The authors stress the limits of the result: a narrow setting, early-2025 tools, and possible learning effects. They also argue that benchmarks and anecdotes each misjudge real-world usefulness.Cited in lesson 1 TalkMost devs don’t understand how context windows workMatt PocockExplains what an LLM context window is (input plus output tokens), why every model has a hard limit, and why performance degrades as context grows (lost in the middle). Shows how to monitor context in Claude Code, when to /clear vs /compact, and warns that MCP servers and large rules files bloat context. Main argument: the most common skill issue among devs is not managing the context window.Cited in lesson 4 TalkOpenAI Codex Masterclass — Vaibhav Srivastav & Katia Gil GuzmanVaibhav Srivastav (VB), Katia Gil GuzmanA hands-on workshop from OpenAI's DevEx team on Codex, OpenAI's software engineering agent. It covers the model-plus-harness architecture, the Codex app with worktrees, plugins, skills, automations, code review, custom subagents, and experimental features such as guardian approvals and hooks. The main argument is that as you delegate more work to parallel agents, you need automated review, scoped subagents, and reusable bundled workflows.Cited in lesson 13 TalkPrompt Injection, explainedSimon WillisonSimon Willison gives a high-level explanation of prompt injection, an attack on applications built on LLMs where untrusted input overrides developer instructions. He argues that AI-based defenses (prompt begging, input/output filtering with more AI) cannot reach 100% reliability and so fail as security. He proposes the dual LLM (privileged/quarantined) pattern as an imperfect mitigation and urges developers to be willing to say no to unsafe designs.Cited in lesson 21 TalkSandboxing AI agents: from containers to gVisor, Firecracker and Katabase caseA walkthrough of how to sandbox code-executing AI agents, building from chroot, jails and Linux containers up to gVisor (application kernel), Firecracker (microVMs) and Kata Containers (VMs behind Kubernetes). It opens with a reported incident in which models escaped a test environment via a kernel bug and the one allowed outbound service. Main argument: containers share the host kernel, so untrusted agent code needs a stronger wall, and egress and credentials must be locked down too.Cited in lesson 21 TalkSimon Willison: Engineering practices that make coding agents work - The Pragmatic SummitSimon WillisonSimon Willison describes his current agentic workflow (writing most code via Claude on his phone) and the engineering practices that make agents reliable: red/green TDD, manual testing with tools like Showboat, conformance suites, and templates that seed good patterns. He also covers prompt injection, the lethal trifecta and sandboxing, and reflects on the exhausting pace, shifting careers, and open source under pressure from agents.Cited in lesson 24 ArticleSubagents - Agentic Engineering PatternsSimon WillisonA short chapter from Simon Willison's Agentic Engineering Patterns guide explaining subagents as a way to work around context limits. A subagent is a fresh copy of the agent with a new context window, dispatched like a tool call. He walks through Claude Code's Explore subagent, then covers parallel and specialist subagents, and warns against over-splitting tasks into too many specialists.Cited in lesson 15 TalkSubagents & the Multiplication of Context Windows Thorsten BallThorsten BallA short talk arguing that subagents matter because they change our relationship to the context window. Each subagent gets its own fresh context window, so noisy tool calls stay out of the main agent's context and only the final result comes back. Ball demos a search subagent, an 'Oracle' subagent backed by a different model (o3), and parallel subagents that edit files.Cited in lesson 15 TalkThe complete guide to Agent SkillsBurke HollandA hands-on walkthrough of Agent Skills in VS Code. Skills are folders with a SKILL.md (name + description metadata) that can bundle scripts and templates. The agent loads them progressively, so they cost little context. The video shows how to build a hello-world skill, then uses Anthropic's PDF skill to teach the agent a capability it lacked, and ends with guidance on skills vs instructions vs prompt files vs custom agents.Cited in lesson 13 TalkThe Lethal Trifecta Is Already on Your Laptops — Michael Patterson, CoderMichael PattersonA vendor talk on Simon Willison's 'lethal trifecta' (private data access, external communication, and untrusted content) as it applies to coding agents running on developer laptops. Patterson argues agents should run in isolated remote environments, with a model proxy and an agent firewall, and he covers the organizational blockers (CISOs, shadow AI, alert fatigue) to rolling agents out at scale.Cited in lesson 21 TalkThe Model Context Protocol (MCP)Theo, DavidAnthropic's MCP team explain what the Model Context Protocol is (a standard way to give context and tools to LLM applications via tools, resources and prompts), how it originated from copy-paste frustration, grew bottom-up from an internal hackathon, and was open-sourced. They cover the state of the ecosystem, tips for getting started, and the roadmap: registry API, long-running tasks, elicitation, and security primitives.Cited in lesson 12 TalkThe Multi-Agent Architecture That Actually Ships — Luke Alvoeiro, FactoryLuke AlvoeiroLuke presents Factory's 'Missions', a multi-agent system that runs coding work for hours to days using an orchestrator, workers, and validators. He argues the bottleneck is human attention, not model intelligence. Key design choices: a validation contract written before code, fresh-context workers with structured handoffs, serial feature execution with parallel read-only work, and the right model in each role.Cited in lesson 16 TalkThe Truth About Coding Agents: Why 90% of Your Time Is Now Code ReviewBeyang LiuSourcegraph's CTO discusses building the Amp coding agent: agent-centric (not model-centric) design, smart vs fast agents, using small and open-weight models for sub-agents, and why code review is becoming the human bottleneck. The last third turns to US open-weight model policy and the dependence on Chinese open models.Cited in lesson 9 ArticleUsing Git with coding agents - Agentic Engineering PatternsSimon WillisonA short guide on using Git together with coding agents. Agents are fluent in Git, so developers can delegate hard operations (merge conflicts, bisect, reflog recovery, history rewriting, repo extraction) and no longer need to memorize the commands. Git history also works as a way to seed an agent's context and as a safety net for reversing mistakes. The article is partly truncated: several example prompts are missing from the text.Cited in lesson 14 ArticleVibe coding, one year later (post on X)Andrej KarpathyKarpathy reflects on the one-year anniversary of his 'vibe coding' tweet. He argues that vibe coding was a fun, throwaway-project mode when LLMs were weaker, while today professionals increasingly work through LLM agents with more oversight. He proposes 'agentic engineering' as the name for this discipline, which keeps software quality high while claiming the leverage of agents.Cited in lesson 1 ArticleWhat is agentic engineering? - Agentic Engineering PatternsSimon WillisonOpening chapter of Willison's Agentic Engineering Patterns guide. It defines agentic engineering as developing software with coding agents, which he describes as agents that run tools in a loop and can execute code. He argues that humans still own deciding what to build, supplying tools, specifying problems, and verifying results, and he distinguishes this from vibe coding, which should mean unreviewed prototype-quality code.Cited in lesson 1 TalkWhy AI evals are the hottest new skill for product builders | Hamel Husain & Shreya ShankarHamel Husain, Shreya ShankarHamel and Shreya walk through a real error-analysis process for an AI app (Nurture Boss): write open-coded notes on traces, cluster them into axial codes, count them, then build narrow binary LLM judges validated against human labels. They argue evals are mostly data analysis on your LLM app, not just unit tests, and that looking at your data is the highest-ROI activity. They also address the 'evals vs vibes / A/B tests' debate and common misconceptions.Cited in lesson 20 ArticleWriting effective tools for AI agents—using AI agentsKen Aizawa (Anthropic)Anthropic's engineering guide to designing tools and MCP servers for LLM agents. Tools are a contract between deterministic systems and non-deterministic agents, so they must be designed for agents rather than as thin API wrappers. The recommended loop is prototype, build an evaluation, then analyze transcripts and iterate with Claude Code. Core principles: choose a few well-targeted tools, namespace them, return high-signal and token-efficient responses, and prompt-engineer tool descriptions.Cited in lesson 12 Article12-Factor Agents: patterns for reliable LLM applicationsDex HorthyThe README introduces 12-Factor Agents, a set of engineering principles for building reliable LLM-powered software for production customers. Dex Horthy argues that the "prompt + bag of tools + loop until done" pattern tops out around 70-80% quality, and that good agents are mostly deterministic software with LLM steps placed carefully. He recommends adding small, modular agent concepts to existing products rather than adopting a framework wholesale.ArticleAI should help us produce better code - Agentic Engineering PatternsSimon WillisonWillison argues that coding agents should raise code quality, not lower it. Shipping worse code with agents is a choice. Agents make simple-but-tedious technical debt fixes and exploratory prototyping cheap enough to adopt a zero-tolerance attitude to code smells, and a compound engineering loop lets those quality gains accumulate.TalkAndrej Karpathy — “We’re summoning ghosts, not building animals”Andrej KarpathyKarpathy argues this is the 'decade of agents', not the year, because current LLM agents still lack continual learning, multimodality and real cognitive depth. He describes how he actually uses coding tools (autocomplete over agents for novel code, agents for boilerplate), why RL is a noisy way to train models, and why demo-to-product gaps (the 'march of nines') apply to software as they did to self-driving. He also covers model collapse, the 'cognitive core', AI as continuous with computing and automation, and his education project Eureka Labs.TalkBuilding Codex with Tibo SottiauxTibo SottiauxTibo Sottiaux, who leads Codex at OpenAI, explains how Codex began as internal research agents and why it was built in Rust and open sourced. He also covers how the harness and model co-evolve, how code review and maintenance are being automated, how the Codex and ChatGPT merge worked, and how the team uses Codex for everything internally.TalkBuilding headless automation with Claude Code | Code w/ ClaudeSedaraA Code w/ Claude talk introducing the Claude Code SDK (headless mode via `claude -p`) and the Claude GitHub Action built on top of it. Includes a live demo where @claude on GitHub issues and PRs produces a feature implementation and a color change, plus a walkthrough of SDK flags for permissions, structured output, session resume and MCP-based permission prompts.TalkClaude Fable, Claude Tag, and Anthropic's Culture — Cat Wu & Thariq Shihipar ft Simon WillisonCat Wu, Thariq ShihiparA fireside chat with two Claude Code team members about how their work has changed with newer models (Opus 4.8, 'Fable'), the launch of Claude Tag (a multiplayer, proactive Slack agent), and how Anthropic handles code review, evals, system prompts, tools, and security (auto mode). The main argument is that as models get better you should trust them with more judgment, remove constraints and humans from the loop only after building eval infrastructure, and push your ambition and product taste higher.TalkConductor CEO Charlie Holtz Walks Us Through His AI Coding SetupCharlie HoltzCharlie Holtz walks through his daily AI coding setup: running many Claude Code/Codex agents in parallel via Conductor, voice-driven prompting, and reviewing PRs in-app. He argues humans should remain architects and own the core, keep 'slop-free zones', and treat code as 'sawdust' while prompts are what matter.TalkFrom fork() to Fleet: Designing an Agent Sandbox Cloud — Abhishek Bhardwaj, OpenAIAbhishek BhardwajA first-principles tour of how to run untrusted agent-generated code securely and at scale. It walks the isolation spectrum from fork/exec to containers, gVisor and microVMs, then covers disk persistence (incremental snapshots, checkpoint/restore) and fleet orchestration. His main argument: use microVMs from the start, and treat durable storage as the next unlock for long-running agents.TalkFrom IDEs to AI Agents with Steve YeggeSteve YeggeSteve Yegge lays out his eight levels of AI adoption for engineers, from no AI to orchestrating many agents. He argues that exponential model improvement will reshape the industry: big companies shrink, small teams of 2-20 people rival their output, and AI use drains developers in a 'vampiric' way. He also describes lessons from building the Gas Town orchestrator, including 'heresies' in vibe-coded codebases and the bitter lesson.ArticleHoard things you know how to do - Agentic Engineering PatternsSimon WillisonSimon Willison argues that a core engineering skill is accumulating a personal collection of working, proven solutions to technical problems. In the agentic era this hoard becomes a powerful input for coding agents: tell them to combine existing working examples to build something new. He describes how he stores solutions (blog, TIL, GitHub repos, HTML tools, research repo) and how he points agents at them.ArticleHow Anthropic teams use Claude CodeAnthropic (article authors)Anthropic interviewed its own teams (Infrastructure, Product Engineering, Security Engineering, Inference, Data Science, Product Design, Growth Marketing, Legal) about how they use Claude Code. Use cases span codebase navigation, testing and review, debugging, prototyping, documentation, and automation. The main argument is that agentic coding dissolves the boundary between technical and non-technical work, and works best as a thought partner rather than a code generator.Talkpeople are lying about agentic codingLars FayeLars Faye argues that the industry narrative that 'coding is solved' conflicts with the reality that agents need detailed specs, rigid feedback loops, strong verification and architectural pipelines, all of which demand domain expertise. He says developers who struggle are told 'skill issue', which feels like gaslighting. He advocates using agents as a delegation layer while staying close to the code, optimizing for doing better rather than faster.TalkPrompting for Agents | Code w/ ClaudeHannah, JeremyAnthropic applied AI engineers explain how prompting differs for agents (models using tools in a loop) versus classic structured prompts. They cover when to use agents, thinking like your agent, giving heuristics, tool selection and design, guiding thinking, managing context, and how to evaluate agents with small realistic evals.ArticlePrompts I use - Agentic Engineering PatternsSimon WillisonA short, continually updated section of Simon Willison's Agentic Engineering Patterns guide. It collects the reusable prompts he actually uses: Claude Artifacts for small HTML tools, a proofreader, alt text drafting, and podcast highlight extraction. The prompt text itself is not in this extract, but the framing shows a practical pattern: saved custom instructions in Claude Projects, plus a firm line on what he won't delegate to an LLM.TalkThe Future of MCP — David Soria Parra, AnthropicDavid Soria ParraA state-of-MCP talk covering MCP's growth (110M monthly downloads), the claim that 2026 is about putting general agents into production where connectivity is key, and a framework of skills, MCP, and CLI/computer use as complementary tools. He urges client-side progressive discovery and programmatic tool calling, designing servers for agents rather than wrapping REST APIs, and previews upcoming spec work (stateless transport, async tasks, cross-app access, server discovery, skills over MCP).TalkVibe coding in prod | Code w/ ClaudeEricEric argues that as AI task length doubles roughly every seven months, engineers must learn to 'vibe code in prod' responsibly, forgetting the code exists but not the product. The way to do it safely is to act as Claude's product manager, confine vibe coding to leaf nodes, and design systems so correctness can be verified without reading the code. He backs this with a 22,000-line change merged into Anthropic's production RL codebase.ArticleWriting about Agentic Engineering PatternsSimon WillisonSimon Willison announces Agentic Engineering Patterns, a guide of evergreen, chapter-shaped patterns for getting good results from coding agents like Claude Code and Codex. He defines agentic engineering as professionals using code-executing agents to amplify existing expertise, in contrast to vibe coding. The first two chapters cover the collapsing cost of writing code and red/green TDD.ArticleWriting code is cheap now - Agentic Engineering PatternsSimon WillisonWillison argues that the central shift in agentic engineering is that writing code is now cheap, which upends engineering habits built around code being expensive. Delivering *good* code (verified, tested, simple, documented, well-designed) still costs real effort, so developers need new habits, such as firing off speculative prompts in async agent sessions instead of dismissing ideas as not worth the time.