Better with Kent · draft

Verification is what closes the loop

Not the diff. The result.

Hook

The code looked fine

The update text was cut off in the window

Nobody ran it. That’s a missing Observe. Theo

Theo · Boris

Coding is solved

Bugs are not

Typing in the editor ≠ idea → software people can use. Theo’s unpack

Lauren Tan

Verification is the skill

The most important one in the toolbox

Same week. Same missing piece. Lauren’s talk

The loop

Act got cheap

Trigger → Act → Observe → Stop

PR + green CI is Act — not closed. ep 05 · the loop

Bottleneck

You are the verifier

  • You open the app
  • You paste the screenshot
  • You cannot parallelize that

No trust → micromanage one agent. Lauren

Definition

Run it like a user

  • Browser · CLI · simulator
  • Trace · heap · preview URL
  • Correct first — good is later

Hardness test

If you can’t check it

the agent can’t either

Spinning up harder than the change → everyone skips it. Theo

Receipt

Every PR is a URL

kody-pr-N.kody.workers.dev 

npm run preview:manual-test -- \
  --request 'POST /onboarding/checklist-dismiss.json {}' \
  --request 'GET /onboarding.json' \
  --check /onboarding

Isolated workers + a seeded login. Health smoke is not the test. preview-manual-testing.md · preview-manual-test skill · full command

Then use it

The written UI pass

  • Create data for this change
  • Open the preview · sign in
  • Click the path · stay on that origin
  • Write what you saw on the PR

Cloud Agents: computerUse. preview-manual-testing.md · preview-manual-test skill

Cheap Observe · 1

Teach the agent the product

  • Feature map: how to get there PStack
  • Vague “???” screenshots become paths
  • DOM hooks · shortcuts · user names

Cheap Observe · 2

Hard gates beat comments

  • Ban the footgun in CI
  • Import boundaries that fail red
  • Soft rules get forgotten

Shortest path should be the best path. PStack · Lauren

Smell

The same PR comment again

Turn it into a lint

Human review as the only invariant is the anti-pattern. Lauren

Cursor

Observe got cheaper

  • Browser / CDP in the loop
  • Cloud agents with a real machine
  • Bugbot · automations · ship-pr

Cheaper ≠ automatic

This week

The fleet already has sensors

That last one is the failure mode

Three surfaces

CI · browser · prod

Tests / lint → Use the UI → Error gone?

Green checks are one surface — not the loop. ep 13 · ship-pr

Trust curve

Don’t spawn 100 yet

  • Watch one until Observe is cheap
  • Then cloud · then automations
  • Auto-merge is earned

Lauren’s trust curve

Homework

One unused change

  • Pick something you never actually used
  • Write the smallest agent check
  • Or: one recurring comment → CI

Better with Kent

Close it through verify

A failure should start the next turn

Like · comment · subscribe · get better with me