My agentic workflow: worktrees, TDD, reviews and a master

· Published · 2 min read

You hear a lot about AI… but rarely anything concrete. So here goes: this is my current agentic workflow.

First, the setup

  • Monorepo codebases, with the ability to spin up a complete instance (e.g. API + DB + user client + admin client + app + containers for every external service, often faked locally) isolated per worktree.
  • Orca, with 5-6 sessions in parallel on average.
  • 1 worktree = 1 complete task, from A to Z.
  • An advanced harness (and I don’t just mean Claude Code):
    • lots and lots of custom ESLint rules;
    • tsc (TypeScript);
    • batteries of tests;
    • a mandatory loop that forces every agent to do everything in TDD. Red, Green, Refactor => careful, not test first. Real mini-iterations;
    • all of it triggered at the right moment with hooks, for a good balance between performance and guardrails.
  • A review system with 8 agents, each specialised in one topic (security, performance, following conventions, DDD… etc). Each one runs 5x with a fresh context, to stay unbiased on every round, with a cross-check system for false positives. And at the end of the review: a very thorough artifact presenting the work done, with diagrams and a change-by-change explanation, the way a colleague would.

This is the kind of setup I help teams build on their own codebases.

Then, my day

  • I scope 4-5 topics.
  • I start a worktree with its own session for each.
  • I start a “master” session, whose job is to check every ~1 min whether one of the others is stuck.
  • As soon as an agent thinks it’s done => the master kicks off the review.
  • Review done: if things were still found in the last rounds => run it again.

And the master tells me when something is 100% done: implementation + review round(s) + artifact.

That’s when I step in for the final check… often 95% perfect.

And while the ~5 agents run, I think and scope the next topics, or I learn things.

Day-to-day work in this job has nothing in common with what it was just 9 months ago…

Happy building!

PS: I’m on Claude Max at $200… and I still hit the weekly limits. That’s not a goal in itself (token maxing is silly, I think), but TDD cycles + reviews are hungry. I’d rather go a bit slower with (over-)checked work than have to put up with nonsense all day long.

PPS: Opus 5.5 on High effort almost all the time.

PPPS: oh! And the upside of the master is that with a quick /remote-control I can check on all of this from my phone. Handy, especially during the sometimes long waiting phases.

Sources