Dev.to AI 🤖 Ai 👁 0 📖 3 min read

Live Coding with an AI Team: A Safe Way to Review a PR (One Writer, One Reviewer, One Publisher)

I ran a small pull request review live with AI agents, and the part that went wrong was never the code. It was the handoff: two agents editing one folder, a bot "reply" that proved nothing, a caption that said "checks pa

I ran a small pull request review live with AI agents, and the part that went wrong was never the code. It was the handoff: two agents editing one folder, a bot "reply" that proved nothing, a caption that said "checks passed" while CI was still queued.

So I turned the setup into a rule you can follow on stream or at your desk:

One issue. One independent check. A human decides what ships.

Below is the short version. The full free guide (41 pages, diagrams, prompts, worksheets) is linked at the end.

1. Give every tool one job

Assign a job before you assign a model.

Tool Its one job It must NOT
Claude Code writes the test and the small fix push or merge
OpenCode reviews the same diff, no edits edit files
Hermes (optional) summarises the handoff decide what ships
ChatGPT dot narrates facts you confirmed announce unconfirmed results

If two agents both own the fix, you get two conflicting edits or two reviews nobody closes.

2. One writer, one folder

A Git worktree is another checkout of the same repo. Give the writer its own and make sure it starts clean:

git worktree add ../demo-fix -b demo/regression-test main
cd ../demo-fix
git status --short   # must be empty before the agent starts

Never run one agent in your main checkout and another in the same folder.

3. Review the exact thing you will publish

A reviewer looking at a stale branch writes plausible comments about code you never wrote. Pin the base and the target, and hand over a saved diff:

git rev-parse --short main            # base SHA
git rev-parse --short HEAD            # target SHA
git diff main...HEAD > review.diff    # exact diff for the reviewer

Then ask for evidence, not opinions:

Review this exact diff against the base. Do not change files.
Tell me what can fail, where, and whether you reproduced it.

A finding is only worth acting on once it has climbed from suspected to reproduced.

4. Prove one connection at a time

Slack, Orca bots, Hermes and dot are separate surfaces. A connection on one grants nothing on the next:

  • A Slack mention does not give repo access.
  • A native Orca group handoff does not prove Slack bot-to-bot delivery.
  • A visible bot message does not prove another bot received a trigger.

Test one human mention, one reply, then stop. Change one variable at a time when something is silent.

5. Allow exactly one publisher

Duplicate PR comments come from repeated triggers or several bots posting the same result. Keep early reviews in local notes or one Slack thread, and give "post to GitHub" to a single identity tied to a specific commit. If a duplicate appears, pause the publisher and save the links before you delete anything.

6. Say the exact CI state

A passing PR check describes the code tested in that check. After the merge, look at the workflow for the merge commit on main.

You may say Only when
queued the run exists but has not started
running jobs are executing
passed the run for that commit finished green
failed any required job is red

If you use a voice agent to narrate, give it one fact, one caveat and one question, and tell it not to announce anything you have not confirmed.

7. If you stream it

Capture one window, not the whole desktop, keep a Break scene ready before any sign-in dialog, and record 30 seconds to check the audio before you go live. Tokens, notifications and private repo names should never be on screen.

Get the full guide (free)

I packed this into a 41-page visual PDF: 5 units and 20 concepts, a prompt pack, a "when things go wrong" debug card, a pre-live checklist and a run-of-show worksheet. It is free:

https://himanshu231204.gumroad.com/l/live-coding-with-ai-team

I build and explain this kind of AI agent workflow live on YouTube: https://www.youtube.com/@BitYodha*Independent community guide. Tool features change, so check each official doc before relying on a setting.*

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.