AI Workshop
Episode 4 of 10Aired

Steering the agent

Plan → execute → verify. Targeted edits instead of rerolls, and the honest moment when it's time to stop.

Wednesday 28 October at 19:00 Moscow time· 90 min

Aired

The recording is being prepared — it usually appears within two days. Sign up for the next streams so you don't miss them.

What we'll show

  1. Plan mode: the agent proposes steps, you edit the plan before it touches code
  2. Executing in small steps with a check after each
  3. Targeted edits: naming the file, the spot and the change so the agent doesn't rewrite more than needed
  4. Rolling back: checkpoints, git and getting back to the last working state
  5. Signs the agent is going in circles, and what to do instead of «try again»

What you take away

A working «plan → code → verify» rhythm and the habit of stopping the agent in time.

Checklist: an agent under control

  • Yes: For any task touching more than one file I asked for a plan first and read it
  • Yes: The plan names the files to change and how the result will be verified
  • Yes: Every step ends with a check: typecheck, a test, the page open in a browser
  • Yes: Edits are targeted: file → spot → change → what not to touch
  • Yes: Before a risky step there's a commit or checkpoint to return to
  • Yes: Three failed attempts in a row — stop: narrow the task, add context or take the step myself

Harness in this episode

Steering is the manual layer of the harness: a plan and a check after each step catch mistakes while they cost minutes, not days.

The full harness episode

What to have ready

  • Yes: Setup and project from the previous episodes
  • Yes: A task you've been putting off

Official documentation