// Generated by scripts/build.mjs. Do not edit by hand. export type Asset = { type: string; text?: string; base64?: string } export const slideCount = 60 export const assets: Record = { "index.html": { "type": "text/html; charset=utf-8", "text": "\n\n\n\n\n\n\n\n\nClosing the Agentic Loop · Kent C. Dodds · BIOS SPHERE 2026\n\n\n\n\n\n\n\n\n\n\n\n\n
\n
\n\n\"\"\"\"\n

Closing the Agentic Loop

\n
Kent C. Dodds
\n
BIOS SPHERE 2026
\n
\n\t
\"\"
\n\t
\n
\n\n
\n
\n\n

The software factory at speed

\n
1,306
merged PRs
91%
from cloud agents
40 min
median, open → merge
\n

Last 30 days. Solo* open-source project: kody-hq/kody

\n

Three pillars make this safe, not just fast.

\n\n
\n
\n\n

Lineage

\n\n\n
\n
\n\n

What Kody is

\n
    \n\t
  • A third place for memory, automations, credentials
  • \n\t
  • Any agent can use it. No inference.
  • \n\t
  • Agents use credentials but never see them
  • \n
\n

Optional: everything today works without it.

\n\n
\n
\n\n

The loop

\n
\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\n\t
Build
Verify
Deploy
\n\t
Outcomes
you want
\n\t
Signal
ideas · feedback
bugs · friction
\n
\n\n
\n
\n\n

Earn the trust to let agents ship

\n
\n\t
1Good PRs
\n\t
2Review + verification
\n\t
3Limit the blast radius
\n
\n

Then the job changes: product + outcomes

\n\n
\n
\n\n

Setup and preflight

\n
\n\t
\n\t
    \n\t\t
  1. Clone your repo, open it in your agent
  2. \n\t\t
  3. Agent CLI and keys ready, Node 20+
  4. \n\t\t
  5. No repo? Fork kentcdodds/mediarss
  6. \n\t
\n\t

Caveat: I've only run this on TypeScript so far.

\n\t
\n\t

Setup

\n
\n\n
\n
\n
1
\n
\n
Good PRs
\n

Make every change checkable

\n\n
\n
\n
Run autopilot
\n

npx kcd-skills@latest autopilot

\n\n
\n
\n\n

Your turn

\n

Tell your agent to run, in a repo you own:

\n
npx kcd-skills@latest autopilot
\n
    \n\t
  • No file changes until you approve a plan
  • \n\t
  • Kody is optional
  • \n
\n\n
\n
\n\n

The autopilot ladder: 13 rungs

\n
\n\t
P1 · Good PRs
1Agent map + verify
2Decisions directory
3Primitives catalog
4Self-verification
5PR template
\n\t
P2 · Review and verification
6Ship loop
7Reviewers + previews
P3 · Blast radius
8Flags + rollback
9Backups
\n\t
KI · Keep improving
10Friction form
11Triage
12Cleanup issues
13Cleanup cron agents
\n
\n\n
\n
\n\n

What autopilot does

\n
    \n\t
  • Audits your repo against the 13 rungs, read only
  • \n\t
  • You approve a plan, then merge one setup PR
  • \n\t
  • Agents take the top priority, prove it, ship it
  • \n\t
  • Product calls and one-way doors wait for you
  • \n
\n

Your exercise: in a repo you own, tell your agent to run

\n

npx kcd-skills@latest autopilot

\n\n
\n
\n\n

AGENTS.md is a router

\n
    \n\t
  • Small: what it is, how to verify, a docs map
  • \n\t
  • A static check fails when it grows too big
  • \n\t
  • As little instruction as possible
  • \n
\n

kody-hq/kody · AGENTS.md

\n\n
\n
\n\n

Primitives: what exists, where it lives

\n
Entry pointsIdentity + authAssistantRuntimeStorage
\n
\n\tFeature flagspackages/worker/src/feature-flags/\n\tScheduled jobspackages/worker/src/jobs/\n\tEmailpackages/worker/src/email/\n
\n

Building blocks, not features. Kept short.

\n

kody-hq/kody · primitives.yaml

\n\n
\n
\n\n

Decision records

\n
    \n\t
  • Context, decision, consequences
  • \n\t
  • Each one says when to revisit it
  • \n\t
  • Changed your mind? Mark it superseded
  • \n
\n

kody-hq/kody · docs/contributing/decisions

\n\n
\n
\n

Friction becomes static checks

\n
    \n\t
  • As static as possible: types, lint, tests
  • \n\t
  • Agents add a rule when something cost them time
  • \n\t
  • Only when you can't, bring in an LLM
  • \n
\n\n
\n
\n\n

The PR template is the rubric

\n
\n\t
Intent
what it's for
\n\t
Door
one-way or two-way
\n\t
Cleanup
none, needed, or done
\n\t
System recap
primitives touched, risk
\n
\n

kody-hq/kody · pull_request_template.md

\n\n
\n
\n\n

One-way vs two-way doors

\n
\n\t
Two-way

Cheap to reverse. Ship it, change it if wrong.

\n\t
One-way
  • First + last name into one field
  • Public API contracts
  • Deleting user data
\n
\n

Feature flags turn a one-way door into a two-way door.

\n\n
\n
\n

Lab 1: Kick off autopilot

\n
\n\t
    \n\t\t
  1. Tell your agent to run npx kcd-skills@latest autopilot
  2. \n\t\t
  3. Check the audit started
  4. \n\t\t
  5. Go to lunch. It runs while you eat.
  6. \n\t
\n\t

Back in

\n
\n\n
\n
\n
Questions?
\n

Next: Review + verification

\n

Common questions

\n\n
\n
\n

Lab 2: Review what your agent produced

\n
\n\t
    \n\t\t
  1. Read the report: scores and evidence
  2. \n\t\t
  3. Answer the interview, read the plan
  4. \n\t\t
  5. Pick one gap to engineer in Lab 3
  6. \n\t
\n\t

Lab 2

\n
\n\n
\n
\n
2
\n
\n
Review + verification
\n

Earn trust for agent merges

\n\n
\n
\n\n

ship-pr: one tick at a time

\n
\n\t
Ready
→
\n\t
CI
→
\n\t
AI review
→
\n\t
Merge or park
\n\t
↓
\n\t
Deploy + health
←
\n\t
Friction
←
\n\t
Report
\n\t
↻ tick until done
\n
\n

Each tick reads where the PR stands. The agent decides.

\n\n
\n
\n\n

~9 taps. No error. They left.

\n
\n
    \n\t
  1. Sentry replay
  2. \n\t
  3. 1:21Agent opens a PR
  4. \n\t
  5. 1:24CI ✗
  6. \n\t
  7. 1:50CI ✓
  8. \n\t
  9. 3 AI reviewers · 8 findings
  10. \n\t
  11. 2:148 fixes
  12. \n\t
  13. 2:25Merged by kody-bot
  14. \n
\n
\n\t
\"QR
\n\t
kody-hq/kody#3782
\n
\n
\n\n
\n
\n\n

It showed its work

\n
\n\t
Before
\"Before:
\n\t
After
\"After:
\n
\n

kody-hq/kody#3765

\n\n
\n
\n

Agents need a way to see the app

\n
    \n\t
  • One CLI to run and check it locally
  • \n\t
  • One skill that teaches it
  • \n\t
  • No throwaway scripts
  • \n
\n\n
\n
\n\n

control-kody

\n
\n\t
\n\t\t
Prove it locally
\n\t\t
dev · login · request --contains
\n\t\t
execute · browse --record
\n\t\t
preview --pr (mocks, seeded user)
\n\t\t
after deploy: health --sha
\n\t
\n\t
\n\t\t
\"QR
\n\t\t
kody-hq/kody
\n\t
\n
\n\n
\n
\n\n
Feature tests prove features.
Health checks prove the running version.
\n\n
\n
\n
Questions?
\n

Next: Break, then the blast radius

\n

Common questions

\n\n
\n
\n

Break

\n\n

Next: limit the blast radius

\n\n
\n
\n
3
\n
\n
Limit the blast radius
\n

Make mistakes cheap

\n\n
\n
\n\n

Risky changes ship dark

\n

Every flag gets:

\n
\n\t
Owner
\n\t
Metric
\n\t
Removal condition
\n
\n\n
\n
\n

Rollback: rehearse it, don't discover it

\n
    \n\t
  • Freeze the deploy queue first
  • \n\t
  • Roll back the version, or fix forward
  • \n
\n\n
\n
\n\n

A backup you've never restored is a hope

\n
    \n\t
  • Scheduled
  • \n\t
  • Restore-tested in isolation
  • \n\t
  • Check the data
  • \n
\n

\"If you haven't actually tested it then it's just theater.\"

\n\n
\n
\n

Credentials agents use but never see

\n
    \n\t
  • Code runs in Kody; secrets never come back
  • \n\t
  • Lock a credential to one package
  • \n\t
  • Profiles scope what each agent can reach
  • \n
\n\n
\n
\n\n

Make mistakes cheap

\n
    \n\t
  • Every flag gets an owner, a metric, a removal condition
  • \n\t
  • Backups get restore-tested in isolation
  • \n\t
  • PocketOS: an agent deleted production in 9 seconds
  • \n
\n

\"If you haven't actually tested it then it's just theater.\"

\n\n
\n
\n
Questions?
\n

Next: Loops + the factory

\n

Common questions

\n\n
\n
\n
4
\n
\n
Keep improving
\n

Then the job changes

\n\n
\n
\n\n

The friction loop

\n
\n\t
1Every PR files what slowed it down
\n\t
↓
\n\t
2An agent fixes it
\n\t
↓
\n\t
3Repeated review comments become checks
\n
\n

friction · P1Invite link rows render \"Expires [object Object]\"

\n\n
\n
\n\n

The signal loop

\n
\n\t
\n\t\t
Errors
\n\t\t
User feedback
\n\t\t
Package errors
\n\t
\n\t
→
\n\t
Gatesignature, rate check
\n\t
→
\n\t
Agent triages
\n\t
→
\n\t
\n\t\t
Bug: agent opens a PR
\n\t\t
Product call: options for me
\n\t
\n
\n\n
\n
\n\n

How a product decision happens

\n
\n\t
1Talk it through, often by voice
\n\t
2The agent lists open questions, with options
\n\t
3I pick. Agents fill in the rest.
\n
\n

\"Just tell me what you're not confident about.\"

\n\n
\n
\n\n
\n\t
\n\t\t

@kentcdodds/
codebase-cleanup

\n\t\t

Cron agents keep the code, tests, and docs lean. Fork mine.

\n\t
\n\t
\"QR
\n
\n\n
\n
\n\n

Where Kody MCP fits in the loop

\n
\n\t
Searchable capabilities
\n\t
Sandboxed execution
\n\t
Memory
\n\t
Verify before handoff
\n
\n

No Kody? The skill still works.

\n\n
\n
\n\n

Lab 3: Engineer one gap

\n
\n\t
\n\t\t

Pick one:

\n\t\t
Verification checkADR or invariantAI reviewerRisk policyPreview envHealth endpointFriction log
\n\t\t
    \n\t\t\t
  1. Ship it as one PR
  2. \n\t\t\t
  3. Prove it works
  4. \n\t\t
\n\t
\n\t

Lab 3

\n
\n\n
\n
\n

Lab 4: Rerun and compare

\n
\n\t
    \n\t\t
  1. Rerun npx kcd-skills@latest autopilot
  2. \n\t\t
  3. Compare with this morning's report
  4. \n\t\t
  5. Note what moved, and what's next
  6. \n\t
\n\t

Lab 4

\n
\n\n
\n
\n

Show and tell, and close

\n
\n\t
    \n\t\t
  1. The gap you closed, and the proof
  2. \n\t\t
  3. One surprise from your agent
  4. \n\t\t
  5. What's still manual? Hand it to an agent.
  6. \n\t
\n\t

Show and tell

\n
\n\n
\n
\n\n
\n\t
\n\t\t
Thank you
\n\t\t
\n\t\t\t\"\"\"\"\n\t\t\t
Kent C. Dodds
@kentcdodds
\n\t\t
\n\t\t

kentcdodds.com·kody.codes·FAQ

\n\t\t

npx kcd-skills@latest autopilot

\n\t
\n\t
\n\t\t
\"QR
\n\t\t
these slides
\n\t
\n
\n\n
\n
\n\n

Common questions

\n\n

← Back to the talk

\n\n
\n
\n

What does this cost?

\n
    \n\t
  • \"Maybe $20,000, $25,000 in tokens\" so far, maybe more
  • \n\t
  • Versus \"three teams of engineers, like five to seven engineers each\"
  • \n\t
  • \"It's not cheap.\" Free tokens are running out, so I steward them.
  • \n
\n

← All questions

\n\n
\n
\n

You don't read the code. What about bloat?

\n
    \n\t
  • \"Bloat is when there's more code than necessary,\" not a line count
  • \n\t
  • A problem \"in so far as it makes it hard for agents to make changes\"
  • \n\t
  • \"Outcomes matter a lot more than output.\"
  • \n
\n

← All questions

\n\n
\n
\n

Who writes the tests?

\n
    \n\t
  • \"The agents are writing the tests.\"
  • \n\t
  • A testing-principles doc, with examples
  • \n\t
  • Every night, an agent deletes useless assertions from the last 24 hours
  • \n
\n

← All questions

\n\n
\n
\n

LLM review or static checks?

\n
    \n\t
  • \"A funnel\": \"as static as possible\"
  • \n\t
  • \"Only when you can't do that do you bring in an LLM\"
  • \n\t
  • Buy AI reviewers: \"they have thought about this problem a lot more than I have\"
  • \n
\n

← All questions

\n\n
\n
\n

How do agents know which primitives to use?

\n
    \n\t
  • \"I did not craft it. I instructed it to be crafted.\"
  • \n\t
  • The PR's system recap grades reuse of existing primitives
  • \n\t
  • Combining them takes \"a product mind and a vision\"
  • \n
\n

← All questions

\n\n
\n
\n

ADRs: tickets, stale ones, monorepos

\n
    \n\t
  • \"A ticket is... the idea. The ADR is we made this architectural decision and we committed to it.\"
  • \n\t
  • Superseded status sends agents to the current record
  • \n\t
  • Many projects in one repo? ADRs per project.
  • \n
\n

← All questions

\n\n
\n
\n

Is it secure?

\n
    \n\t
  • Secrets \"live in Kody and Kody will never send those secrets back\"
  • \n\t
  • A locked credential does only \"the deterministic thing that your package supports\"
  • \n\t
  • Webhooks check signatures and rates first: no \"denial of wallet\"
  • \n
\n

← All questions

\n\n
\n
\n

What is Kody, and do I need it?

\n
    \n\t
  • \"A modular shared layer that all of your agents have access to... not a harness\"
  • \n\t
  • \"Kody actually does not have any inference.\" Use any agent.
  • \n\t
  • No: \"You can make scripts for all this stuff too.\"
  • \n
\n

← All questions

\n\n
\n
\n

How do you prompt and spec?

\n
    \n\t
  • \"Talk to the agent like it's a person.\"
  • \n\t
  • Ask: \"tell me what you think I told you,\" or \"what would you do if I said go?\"
  • \n\t
  • \"Prompting is really just alignment of ideas.\"
  • \n
\n

← All questions

\n\n
\n
\n

How much do you review? What merges itself?

\n
    \n\t
  • Each repo has a risk policy: ship-pr merges, or parks for me
  • \n\t
  • I look at \"maybe a quarter\" of PRs; big or visual changes wait
  • \n\t
  • Small things sometimes go \"straight on main\"
  • \n
\n

← All questions

\n\n
\n
\n

Product work: design, discovery, priorities

\n
    \n\t
  • Agents help with \"market research,\" feature usage, pricing. \"Talk to tons of customers.\"
  • \n\t
  • One idea at a time; agents work on many things in parallel
  • \n\t
  • \"I am the bottleneck on that and I think that should probably remain.\"
  • \n
\n

← All questions

\n\n
\n
\n

Don't you drown in agent output?

\n
    \n\t
  • \"I mostly don't read agent's output.\"
  • \n\t
  • I read the conversations, often by voice on a walk
  • \n\t
  • \"Just tell me what you're not confident about.\"
  • \n
\n

← All questions

\n\n
\n
\n\n\n\n\n\n\n\n\n" }, "favicon.ico": { "type": "image/x-icon", "base64": "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" } }