Skip to content
← Public packages

@kentcdodds/closing-the-agentic-loop

Closing the Agentic Loop: Kent C. Dodds's talk and workshop deck. Default short talk; ?event=aie-2026 is the AIE hour cut with a Podium QR; ?format=full adds labs.

FollowForkOpen app

Files

LAB-GUIDE.md

132 lines · 6.9 KB · Markdown

Closing the Agentic Loop: full-day lab guide

This is the attendee guide for the full-day workshop version of "Closing the Agentic Loop" by Kent C. Dodds. The slides for the day are this deck with ?format=full (for example https://closing-the-agentic-loop--kentcdodds.kody.run/?format=full). The default URL is the short talk and has no lab slides.

Through the day you run the autopilot skill on a repo you own, review what it finds, engineer one gap yourself, and rerun it to see what moved. Each lab slide has a countdown; the times below are the defaults and the presenter may change them.

BlockDefault timeSlide
Setup and preflight30 min#/lab-setup
Lab 1: kick off autopilot (runs over lunch)75 min, lunch included#/lab-1
Lab 2: review what your agent produced45 min#/lab-2
Lab 3: engineer one gap90 min#/lab-3
Lab 4: rerun and compare45 min#/lab-4
Show and tell, and close30 min#/show-and-tell

Setup and preflight

Do this before the workshop if you can. On stage, the most common failure was a repo that wasn't cloned yet.

  1. Clone your repo and open it in your agent. Pick one you can push branches to and open pull requests on. A small or medium app is easier than a huge monorepo.
  2. An agent that can run commands in the repo, signed in with a subscription or API key: Claude Code, Codex, Cursor, Devin, or similar. Check that it can run a shell command in the repo before you arrive.
  3. Node 20 or newer. The skill runs with npx kcd-skills@latest.
  4. Your code host CLI, if you have one. On GitHub, gh auth login helps: the audit gathers host facts (labels, PR template, review bots) read-only. Any code host works; when the skill needs a host fact it prints what to gather and your agent supplies it.
  5. No repo of your own? Fork kentcdodds/mediarss (public, TypeScript) and work on your fork. Don't open pull requests against the original.

Caveat: Kent has only run this on TypeScript repos so far. The skill is written to work on any language and stack. If yours is something else, you're the experiment; please share what happened at show and tell.

Kody is optional. The interview asks "Use Kody?"; choose no and the skill sets up the same ship loop with npx kcd-skills@latest autopilot ship-pr, an issue form, and CI cron instead.

Lab 1: kick off autopilot

Timed to run over lunch, because the audit runs your repo's verify command and that can take a while.

  1. In your repo, tell your agent: run npx kcd-skills@latest autopilot and follow the instructions it prints.
  2. Check that the audit started: your agent should be gathering facts and running the verify command. If it fails right away, fix setup now (Node version, missing dependencies, host sign-in) and start again.
  3. Go to lunch.

What it does: the audit scores each rung of the ladder (done, partial, missing, unknown, or n/a) from evidence in the repo and on the host. It leaves your tracked files unchanged, though it may install dependencies and run the verify command. Then it renders a report, interviews you, and proposes a plan. Nothing is edited until you approve a plan. If the agent is waiting on interview questions when you get back, that's expected.

Lab 2: review what your agent produced

  1. Read the report (autopilot report --html renders it). For each rung, look at the score and the evidence. Unknown means "ask or verify"; it never turns into a PR until it's resolved. If a score looks wrong, tell your agent why and have it check again.
  2. Answer the interview for your own repo and team, then read the plan: a table of proposed PRs, one per rung, in order.
  3. Pick one gap to engineer in Lab 3. Prefer a missing or partial rung you care about and can finish in one small PR. You don't have to approve the whole plan.

Bring questions about your run to the check-in. This is the time set aside for them.

Lab 3: engineer one gap

Choose one. Each should end as one pull request whose Testing section says what was proven.

  • Verification check. Find one thing that cost your agent time and make it static: a lint rule, a type, or a test. Or give agents a way to prove a change locally: a small control CLI command (start the app, sign in as a seed user, request a page and check it contains some text).
  • ADR or invariant. Add a decisions directory with a template (number and title, status: accepted or superseded, date; then context, decision, and consequences, including the condition that would make you revisit it). Write one real record for a decision your repo already lives by. Or write down one invariant and add a check for it.
  • AI reviewer. Install an AI code reviewer on the repo (for example CodeRabbit, Cursor Bugbot, or Devin) and tell agents to request reviews from bots and to evaluate each comment rather than accept it because it's a review.
  • Risk policy. Write down which changes may merge on green CI (low), which wait for an AI review (medium), and which wait for a human (high). One-way doors (stored data shapes, public contracts, deleting user data) belong in high.
  • Preview environment. A per-PR preview with mocked third parties and a seeded user, so an agent can check a change in a production-like environment.
  • Health endpoint. A /health route that reports the deployed commit, plus a check after deploy that fails unless production serves that commit.
  • Friction log. Somewhere agents file what slowed them down: an issue form and a friction label, with a priority. A local script works too; the hard part is credentials. Then tell agents to file friction after every PR.

Tips:

  • If the Lab 2 plan already has a PR for your rung, let your agent build that PR from the plan and review it with the PR template as the rubric (intent, door, cleanup, testing).
  • Keep it small enough to review. Prove it works before you call it done.
  • Kent's packages are public and forkable on Kody if you want a head start: @kentcdodds/ship-pr, @kentcdodds/friction-log, and @kentcdodds/codebase-cleanup.

Lab 4: rerun and compare

  1. Run npx kcd-skills@latest autopilot again (another read-only audit; it takes a while again).
  2. Compare the new report with this morning's. Did the rung you worked on move (missing to partial, partial to done)? If not, ask your agent what evidence it expected.
  3. Write down what moved and what's next: the next rung from the plan, and one thing you still do by hand that an agent could do.

Show and tell, and close

Two or three minutes each:

  • The gap you closed, and how you proved it.
  • One surprise from your agent.
  • What's still manual, and how you'll hand it to an agent.