How Walden Yan actually uses coding agents
Co-founder, Cognition (Devin)
If you take one thing: Daily customer-report Devin, escalates top feature asks. his favorite Devin job is not code at all, it reads every customer report each day and escalates the top requests to engineering.
- Agent
- Devin (web), primary, runs a team of Devins · 2026-03-19
- Model
- Fable, via Devin Fusion · 2026-07-13
- Model
- GPT-5.5 · 2026-04-24
- Review
- Devin Review first, then asks Devin what to look at first · 2026-03-19
- Harness
- Slack channel feeding a team of cloud Devins; stacked PRs · 2026-03-19
“I spend so much on tokens nowadays but have way more trust in my PRs.”
In their words
- “I’ve had a lot of fun pointing Devin to look at everything in a slack channel, split up the tasks automatically, and then start a team of Devin’s to tackle them.”
- “But then I tried it, and on one of my tasks, Fusion w/ Fable was 4x cheaper and gave a better PR.”
- “I find myself really reaching for 5.5 nowadays. Seems like a model built for cloud.”
- “Then when I go to review, I ask it to tell me what I should look at first”
- “I’ve had a lot of fun pointing Devin to look at everything in a slack channel, split up the tasks automatically, and then start a team of Devin’s to tackle them.”
- “My favorite use for Devin isn’t even writing code. Every day, my Devin reads every customer report and escalates all the top feature requests to our eng team. Eyes everywhere.”
Newest source: 2026-08-19. Is this you? Edit this card on GitHub.
OpenClaw, ex-PSPDFKit
Agents: OpenClaw (web), Codex (app). Model: GPT-6.1 Sol.
Review: Codex reviews every commit
One thing to take: $autoreview, before every commit.
Datasette, Django co-creator
Agents: Codex (app), Claude Code (web). Model: GPT-6 Sol, Claude Opus 5.5.
Review: Cross-model review: Claude checks GPT, GPT checks Claude
One thing to take: Showboat, agents demo their own work.
Created Claude Code
Agents: Claude Code (app), Claude Tag. Model: Fable 5.1.
Review: Reads diffs only for critical code
One thing to take: /go, test, /simplify, open PR.
nanochat, coined “vibe coding”
Agents: Claude Code (cli), Codex. Model: Claude Fable 5.
Review: Watches agents like a hawk in an IDE
One thing to take: autoresearch, overnight agent research.
Ghostty, HashiCorp co-founder
Agents: Claude app (app), Codex app (app). Model: Claude Fable, GPT-6 Astra.
Review: Hand-reviews every line; agents validate, he writes all prose
One thing to take: Harness engineering, after every agent mistake.
Ruby on Rails, Omarchy
Agents: Codex (cli), Claude Code (cli), OpenCode (cli). Model: GPT-6.1 Sol, Claude Opus 5.5.
Review: A second model reviews, then he reads the diff
One thing to take: "Review this with claude", adversarial review from Codex.
Flask, Earendil founder
Agents: Pi (cli), Codex. Model: Sol, Fable.
Review: Steers as it goes, reviews when the agent says done
One thing to take: /answer, in place of plan mode.
T3 Stack, T3 Chat, T3 Code
Agents: Claude Code (app), Codex (app). Model: Opus 5.5, Muse Wonder.
Review: Opus asks Wonder for code review; agents merge when confident
One thing to take: T3 Code, his own GUI over every agent.
Redis creator, DwarfStar (ds4)
Agents: Codex (cli), Claude Code (cli), ds4-agent (cli). Model: GPT-6 Astra / GPT-6.1 Sol, Fable.
Review: Still reviews Redis code, calls it mostly pointless
One thing to take: QA_BEFORE_RELEASES.md, an agent QA pass per release.
Harpoon, Vim on stream
Agents: Cursor (ide), 99 (ide). Model: Grok 4.6.
One thing to take: Field guide, agent writes its own lessons.
Gas Town, Beads, Vibe Coding
Agents: Claude Code (cli), Codex. Model: Claude Opus 5.5, Claude Fable 5.1.
Review: Fable reviews every Opus change; he doesn't read code
One thing to take: Seats, 18 long-lived Fable officers.
Photo AI, Nomads.com
Agents: Claude Code (cli). Model: Claude Opus 5.5, Claude Fable 5.1.
Review: Step by step, keeps the agent on a short leash
One thing to take: tm, a tmux session per site.
pi coding agent, libGDX
Agents: Pi (cli). Model: GPT-5.5.
Review: Limits daily output to what he can review
One thing to take: Daily code limit, sized to what he can review.
Amp co-founder, interpreter book
Agents: Amp (web). Model: Claude Opus 5.5, GPT-6 Astra.
Review: Spot checks and architecture, not every line
One thing to take: “Irrefutable proof”, video proof, not line reads.
Superpowers, Keyboardio, K-9 Mail
Agents: Claude Code (cli), Evener, Muse. Model: Claude Opus, Claude Haiku.
Review: Agent sessions do adversarial review
One thing to take: /par, rival reviewers, 5 points.
CEO 2389.ai, CTO Obama 2012
Agents: Claude Code (cli), Codex (app). Model: Claude Fable 5.1, GPT-6 Astra.
Review: One fresh-eyes review pass per change
One thing to take: "fresh eyes" review, a subagent rereads the code.
TDD, Extreme Programming, JUnit
Agents: Claude Code (cli), Codex (cli). Model: Gemini.
Review: Reviews solo, would rather be pairing
One thing to take: TCR skill, red tests revert the code.
Compound Engineering, Cora
Agents: Cursor, Claude Code (cli), Codex. Model: Claude Opus 5.5, Claude Sonnet 5.5.
Review: Parallel review agents, no line-by-line reading
One thing to take: /ce-compound, turns each fix into a lesson.
Total TypeScript, ex-Vercel
Agents: Claude Code (cli). Model: Claude Opus 5.5, Claude Opus 5.
Review: Reviews by merge danger, two-axis /code-review
One thing to take: /grill-me, before every change.
Context, uniprof, Sentry AI lead
Agents: Claude Code (cli), Codex. Model: Claude Opus 5.5, GPT-5.
Review: Reviews the plan and prompts, not the 10,000-line diff
One thing to take: Claudescope, transcript link in every PR.
Claude Code at Anthropic, ex-Chrome
Agents: Claude Code, Codex. Model: Fable.
Review: Careful human review of core paths only
One thing to take: /loop PR triage, on Agent Skills, every day.
Creator of the Ralph loop
Agents: OpenCode (cli), Cursor (web), Loom. Model: GPT-6.1 Sol, GLM / Kimi.
Review: Ships without manual review; stops only for high-risk changes
One thing to take: The Ralph loop, an agent in a bash while loop.
Co-founded Amp and Sourcegraph
Agents: Amp (web). Model: Amp Dial: medium.
Review: Agent reviews first, he scans diffs for red flags
One thing to take: "I'm feeling Pucky" button, starts half their work.
OpenCode co-founder, SST
Agents: OpenCode. Model: Astra, Fable.
Review: Asks a smarter model to review, via subagents
One thing to take: Voice prompting (Hex), instead of typing.
Codex at OpenAI, ex-Cline
Agents: Codex (app). Model: GPT-6 Astra, GPT-5.3-Codex-Spark.
One thing to take: "Chief of staff" thread, spawns every project thread.
Tailwind CSS, Refactoring UI
Agents: Amp (web), Codex (app), Claude Code (web). Model: Claude Sonnet 4.5.
One thing to take: Every tedious OSS task, fed to cloud Claude Code.
Vitest, Slidev, UnoCSS, VueUse
Agents: Claude Code, Cursor (ide). Model: Claude Sonnet 5.5, Claude Opus 5.
Review: Watches the result: Storybook for agent UI, a pulls.review summary on every PR, then the final diff
One thing to take: antfubot, agents ship PRs as a bot.
Ladybird browser, SerenityOS
Agents: Claude Code, Codex. Model: Frontier model.
Review: Adversarial review passes by different models
One thing to take: Welfare checks, on workers across 4 machines.
DTrace, Oxide co-founder and CTO
Agents: Claude Code (cli), ChatGPT. Model: Claude Opus 4.8.
Review: Self-review first: nobody else reviews LLM code until the responsible engineer has
One thing to take: This Old Repo, reviving a 1990s game.
Thoughtworks, harness engineering
Agents: Cursor (ide), Claude Code, Pi (cli). Model: Claude Sonnet, Claude Opus 4.8.
Review: Risk triage: probability, impact and detectability set how much she reads; sensors point her to where
One thing to take: sensors check, a sidecar the agent polls.
Works on GitHub Copilot
Agents: GitHub Copilot (app), Claude Code (web). Model: Luna Max, Claude Opus 5.5.
Review: Rubber duck review by a second model family; he no longer reads the code
One thing to take: A lead agent per project, he just files issues.
ripgrep, Rust regex, Jiff
Agents: Codex (cli). Model: GPT-5.6 Sol.
Review: Flags his LLM PRs for careful review, then simplifies them by hand; Codex review catches bugs too
One thing to take: "I wrote this PR with an LLM", flagged for careful review.
Segment co-founder, ex-Codex
Agents: Codex (cli), Claude Code (cli). Model: Fable 5, GPT-5.6 Sol.
Review: Spot-checks the architecture; Cursor Bugbot and Codex review every PR
One thing to take: /pr-pass, loops until CI and bots pass.
Claude Code product, Anthropic
Agents: Claude Code (app), Claude Tag. Model: Claude Opus 5.5, Claude Fable 5.1.
Review: Claude code review clears the outer layers; code owners review the core
One thing to take: Fresh config per model, reset at every model launch.
AI Engineering author
Agents: Codex, Claude Code. Model: Claude Fable.
Review: A second agent reviews the plan; UI work comes with screenshot evidence on the PR
One thing to take: Cross-provider runner, Codex, Claude spawn each other.
Creator of Phoenix, Phoenix.new
Agents: Claude Code. Model: Claude.
Review: Rules file names the tests to run before claiming a fix
One thing to take: `notify msg`, in CLAUDE.md, not a hook.
exe.dev, Tailscale co-founder
Agents: Shelley (web). Model: OpenAI models.
Review: Reads agent code before it hits serious systems, trimmed by meat
One thing to take: meat, strips diffs to what matters.
ChatPRD founder, How I AI host
Agents: Codex (app), Claude Code, Cursor (web). Model: GPT-6 Astra, Claude Opus 5.5.
Review: Cross-model review: Opus 5.5 reviews GPT's PRs and vice versa; PRs also get human review
One thing to take: LGTM, the PR closer, a bot that clears her PR queue.
danluu.com essays, ex-Centaur
Agents: Codex (cli). Model: GPT-5.6 Sol.
Review: Heavy randomized testing instead of human review
One thing to take: Named-persona review panels, with contrarian personas.
"How to Fix Your Context", Plumb
Agents: Claude Code. Model: Claude Fable, GLM 5.2.
Review: Steps through the decisions Plumb extracts at each commit instead of reading the diff
One thing to take: Plumb, reviews decisions at commit.
HumanLayer CEO, 12-Factor Agents
Agents: HumanLayer (ide), Codex, Claude Code. Model: GPT-5.6 Sol, Claude Fable.
Review: Still reads the code, one 1k-3k PR at a time
One thing to take: /show-me, diagrams, not walls of prose.
CEO of Every
Agents: Codex (app), Claude Code. Model: GPT-6 Sol, Claude Fable.
Review: Vibe codes; doesn't read the codebase
One thing to take: "Want me to compound?", asked after every sweep.
Codex product lead, OpenAI
Agents: Codex (app). Model: Codex default model.
Review: Reads the plan; Codex reviews nearly every PR
One thing to take: "Using a subagent in parallel, do X", from one always-on thread.
PyTorch at Meta, ghstack
Agents: Claude Code (cli), Codex. Model: Claude Opus 5.5, Claude Fable 5.
Review: Reads every line, steers with questions and edits
One thing to take: sl follow / sl adopt, restack parallel agents.
Game Engine Black Books
Agents: Claude Code (ide), Antigravity (ide). Model: .
Review: Reads all the code; LLM criticizes first
One thing to take: "Reload agent.md", when code quality drops.
Go cryptography, age, mkcert
Agents: Shelley (exe.dev) (web), Claude Code (cli). Model: GPT-6 Astra, Claude Opus 4.8.
Review: LLM review before any human review
One thing to take: filippo-agent, agent's own GitHub account.
YC CEO, creator of gstack
Agents: Claude Code (app), Codex (cli). Model: Claude Fable 5.1.
Review: Every branch gets CEO, Codex, DX and eng review passes
One thing to take: /plan-ceo-review, finds the 10-star product.
Ansible, Raspberry Pi YouTube
Agents: Claude (web). Model: Claude Opus 5.5, Claude Fable 5.1.
Review: Reviews in stages, like a tech lead with juniors
One thing to take: AI Disclosure, README names the exact model.
Malleable software, Notion
Agents: Claude Code, Cursor (ide), Codex CLI (cli). Model: Claude Opus 5, Claude Fable 5.
Review: Reads an explainer first, then the full diff; must pass a 5-question quiz before sending code on
One thing to take: /explain-diff, with a quiz before he ships.
tinygrad, comma.ai
Agents: OpenCode. Model: Kimi, Claude Opus.
Review: Reads and understands each line
One thing to take: (kimi) in the PR title, names the model per PR.
Creator of Bun, now at Anthropic
Agents: Claude Code (cli), RoboBun. Model: Claude Fable 5.
Review: Two adversarial Claude reviewers per implementer, then he reads the outputs
One thing to take: Adversarial review, 2 Claudes refute every change.
Symmathesy, Honeycomb
Agents: Claude Code. Model: Claude Opus 5.5, Claude Sonnet 5.5.
Review: A reviewer agent checks each PR against a stated correctness condition; her own reviews fix the system
One thing to take: Capability Owners, standing agents per feature.
Creator of Elixir, Tidewave
Agents: Codex, Tidewave (app). Model: Sol 5.6.
Review: Reviews agent PRs like any other code
One thing to take: Adversarial review agents, on compiler & security work.
fast.ai, Answer.AI, ULMFiT
Agents: Claude Code (cli), Solveit (web), Codex. Model: Claude Opus 4.5.
Review: Small steps he follows one by one; the agent must propose before touching his published libraries
One thing to take: claude-air, nudges after 8 silent rounds.
Instructor creator, OpenAI DX
Agents: Codex (app). Model: GPT-5.5, GPT-5.3-Codex-Spark.
Review: Review skills modelled on colleagues' PR feedback; memory changes reviewed as git diffs
One thing to take: $ultragoal, goal kept in an editable file.
Testing Library, Epic Web
Agents: Grok Bot, Cursor (web). Model: Grok 4.5, Fable.
Review: Trusts gates and AI reviewers, reads the outcome, not the diff
One thing to take: /ship-pr, Bugbot loop, merge, Discord.
Cloudflare Workers, Cap'n Proto
Agents: Claude Code (cli), Cloudflare OS (web). Model: Fable 5.1, Sol.
Review: Sol reviews every commit; he reviews the plan and the code
One thing to take: plans/ in the repo, checked in for code review.
CEO of Obsidian
Agents: Claude Code (cli). Model: .
Review: Recorded test output must be read; fixture must fail before the fix
One thing to take: Fail-first fixtures, before every extraction fix.
ML at SpaceX, ex-Cursor, Vercel
Agents: Grok Bot (app), Cursor (web). Model: Claude Opus 4.5, GPT-5.1-Codex-Max.
Review: Reviews most of the code and tests it; Bugbot reviews every PR
One thing to take: Plan mode, before every long agent run.
Staff Engineer, CTO at Imprint
Agents: Claude Code, Agent Fleet, Codex. Model: GPT 5.4.
Review: Same harness reviews first, context cleared; humans after
One thing to take: /linear-project-loop, loops a project to its goal.
Digital garden, GitHub Next
Agents: Claude Code (web), Codex, OpenCode. Model: Claude Opus 5, Claude Sonnet 4.6.
Review: Watches the diffs, checks code in an IDE
One thing to take: maggie-design, her design taste, as a skill.
rust-analyzer, TigerBeetle
Agents: Claude Code (cli). Model: Claude 3.7 Sonnet.
Review: Owns the structure, runs the code, does not vet the guts
One thing to take: Hand-typed skeletons, Claude fills in the bodies.
Founder, Takeoff AI
Agents: Codex (app), Claude Code (cli). Model: Claude Opus 5.5, GPT-6 Astra.
Review: Approves without reading the code
One thing to take: "You are AGI-pilled.", in every agent's system prompt.
Homebrew maintainer since 2009
Agents: Codex (cli), Claude Code (cli), OpenCode (cli). Model: Fable 5, OpenAI Codex 5.4.
Review: Reviews agent work before pushing it himself; depth depends on the code
One thing to take: AgentIDE, his own vibe-built agent IDE.
Instagram co-founder, Anthropic Labs
Agents: Claude Desktop (app). Model: Claude Opus 5.5, Claude Sonnet 5.5.
Review: Hands over whole goals and has Claude verify, then double-verify
One thing to take: iOS Claude pings Android Claude, to port each feature.
Ex-BuzzFeed data scientist
Agents: Zed Agent (ide), Codex (cli), Claude Code (cli). Model: GPT-6 Astra, GPT-5.6 Luna.
Review: Luna subagents re-check until all satisfied; he watches benchmark diffs
One thing to take: codex exec Luna subagents, 7–12 cheap critics per pass.
RLHF Book, Interconnects
Agents: Claude Code (cli), Codex. Model: Claude Fable 5, GPT 5 Pro.
Review: Lets agents catch their own mistakes, then checks the results himself
One thing to take: /gemini-feedback, Gemini critiques each diagram.
coverage.py
Agents: Claude Code (cli), GitHub Copilot (ide). Model: .
Review: A person owns and reviews every commit, Claude or not
One thing to take: Be introspective, Claude fixes its own rules.
PouchDB, emoji-picker-element
Agents: Claude Code, Codex (cli), Cursor Bugbot (web). Model: Claude Opus 4.7, GPT 5.5.
Review: Three agents hunt bugs in every PR; he triages the ranked report and guides the fixes
One thing to take: code-review-turbo, three-agent PR bug hunt.
Puzzmo, CocoaPods, Danger
Agents: Claude Code (ide), Claude Code on the web (web). Model: Fable.
Review: Reads all Puzzmo code; Claude reviews every PR too
One thing to take: /commit, human writes the headline.
Chrome DevTools, Lighthouse
Agents: Codex (web), GitHub Copilot coding agent (web), Jules (web). Model: Gemini 3 Flash (preview).
Review: Codex bot reviews every PR; he has Copilot fix the findings
One thing to take: paulirish-agent, a GitHub account for agents.
Build an LLM From Scratch author
Agents: Codex, Claude Code, Qwen Code (cli). Model: GPT-5.5, Qwen3.6-35B-A3B.
Review: Wants to see what the agent is doing; writes the code he cares about himself
One thing to take: Audit the agent first, before installing any harness.
CEO of Vercel
Agents: Claude Code. Model: Claude Opus 4.6.
One thing to take: gdp-ts, auth checks as type proofs.
Creator of Node.js and Deno
Agents: Claude Code, Codex. Model: Claude Opus 4.6.
One thing to take: Claw Patrol, firewall between agents and prod.
Writing on AI and big tech
Agents: GitHub Copilot app (app), Copilot CLI (cli). Model: GPT-6-Astra, Claude Opus 5.5.
Review: Skims each attempt for ~30 seconds, rejects most, then fully reviews the one he keeps
One thing to take: 30-second triage, rejects most agent attempts.
Gumroad founder
Agents: Claude Code, Codex. Model: Claude Fable 5, GPT-6.
Review: Does not just hit merge: QA, then asks a colleague to check
One thing to take: Fourth-wall check, enforced in CI on every PR.
Founder of Untangle
Agents: Devin (web), Codex. Model: Claude Opus 5.5, SWE 1.7.
Review: Agents review and land; he watches a test video first
One thing to take: !land playbook, refreshes skills on every PR.
Co-author of the Rust book, jj
Agents: Claude Code (cli), Codex (app). Model: Claude Fable, GPT Sol.
Review: Codex reviews Fable's PRs; humans stay responsible for what they submit
One thing to take: rue-overnight-hunt, bugs filed only with repros.
Latent Space, AI Engineer
Agents: Codex, Claude Code (cli), Devin (web). Model: Claude Opus 4.6.
Review: Devin Review bot comments on each agent PR, often after merge
One thing to take: $cassandra-creator, a standing dissent agent.
Fields Medal mathematician, UCLA
Agents: Claude Code. Model: Claude Opus 5.
Review: Checks results, doesn't read machine code
One thing to take: Five favorable conditions, before handing code to an agent.
Claude Code team, Anthropic
Agents: Claude Code (cli), Claude Tag. Model: Claude Fable 5, Claude Opus 5.5.
Review: Claude writes an HTML report and a quiz on the change; he merges only after passing it
One thing to take: Change quiz, must pass before merging.
FastAPI, Typer, SQLModel
Agents: Codex. Model: GPT-6 Astra.
Review: Reviews and adjusts all AI code by hand
One thing to take: AI Disclaimer, names agent and model in PRs.
Codex & ChatGPT at OpenAI
Agents: Codex (app). Model: GPT-5.6 Sol, GPT-6 Astra.
Review: Reviews what the agent presents; uses a fast model to understand code, never delegates understanding
One thing to take: Random-file bug hunt, automation, several times a day.
Val Town, Placemark, ex-Mapbox
Agents: Claude Code (cli), Pi. Model: Claude Sonnet 5, GLM 5.2.
Review: Careful human review of LLM fixes
One thing to take: Pi on GLM 5.2, as the security bug hunter.
Shopify CEO, qmd
Agents: Claude Code, Pi (cli), Hermes Agent. Model: Claude Opus 5.5, GPT-6 Astra.
Review: Spec suite the agent must pass
One thing to take: qmd, local search, agents too.
DataScript, Fira Code
Agents: Claude Code, Codex. Model: Claude Opus 5.5.
One thing to take: Live nREPL for the agent, instead of app restarts.
Creator of Linux and Git
Agents: Claude. Model: Claude Opus.
Review: Opus writes verification scripts; the AI checks its own output
One thing to take: Pedals in a ring, a bench that tests itself.
Matasano, Latacora, Cryptopals
Agents: Claude Code (cli), Codex. Model: Claude Opus 5.5, GPT-5.5.
Review: Agents push PRs, only he merges
One thing to take: swiftui-app template, copied into every new app.
Clean Code author
Agents: Grok (cli), Codex (cli). Model: Grok.
Review: Does not read agent code: Codex reviews Grok's output, he inspects a UML diagram colored by CRAP and mutation scores
One thing to take: uml-viewer, colored by CRAP + mutation.
Co-founder, Cognition (Devin)
Agents: Devin (web). Model: Fable, GPT-5.5.
Review: Devin Review first, then asks Devin what to look at first
One thing to take: Daily customer-report Devin, escalates top feature asks.
Syntax podcast co-host
Agents: Claude Code (cli), Codex (app), OpenCode (cli). Model: Claude Opus 5.5, Codex models.
Review: Reviews after the fact: tests it works, checks for duplicated messes
One thing to take: vite-plugin-killer-instincts, hands the LLM a kill command.
pandas, Apache Arrow
Agents: Claude Code (cli), Codex (app). Model: Claude Fable, GPT-5.6 Sol.
Review: roborev reviews every commit; he checks structure
One thing to take: roborev, reviews every agent commit.
Anubis, xeiaso.net
Agents: Claude Code (cli), Codex, OpenCode. Model: Claude Opus 5.5, GPT-6.1 Sol.
Review: A second agent team reviews; /reviewbot on PRs
One thing to take: Assisted-by: footer, model + tool on every commit.
Co-founder and CEO of Warp
Agents: Warp Factories (web), Warp agent (app). Model: GPT 5.6 Sol.
Review: Engineer reviews the agent PR and ships it; QA video attached
One thing to take: Scorer + self-improvement agents, agents that fix agents.
