What made it work for me:

1. One task per session, each in its own git worktree, so two agents never edit the same checkout.

2. Stop polling. Claude Code's Notification hook fires when it's waiting on a permission prompt or sitting idle; point it at notify-send or a desktop toast and you only look at a session when it asks. The hooks guide has the snippet.

3. Given the work on worktrees, I always have PRs which I take a look myself. You can utilize other AI to review them too, which I find useful, just for smaller projects looking at them myself is enough.

first I jot down my notes then I work with one agent to write a plan. this is just the first draft.

Once I have a draft, I start to leave comments to ask questions and launch multiple agents to research and refine parts of the plan.

(optionally share the plan with my team for feedback)

then have multiple agents work on implementing parts of the plan

Ref is great for this because it can orchestrate agents and you can track all the progress on one place. It's hosted so you just work from a browser like Google Docs but can connect to local agents too

1.When I try agent based orchestration and multi tasking, some parts work well overall, but certain areas keep repeating local issues.

2.So I tried having AI review each other's work and set up guardrails, like stopping when certain values or patterns repeat. That didn't work well either.

3.In the end, most of the time, if I define what the 'result' should look like and run it, the system eventually reaches that outcome.

The problem for me is that the orchestration code itself doesn't fit the style I like, so it's hard to look at. Using orchestration code basically means you're trusting the AI to handle everything without looking at the code. A lot of people say results come out quickly and work well, and models around the SOL level seem to perform decently. But for me, I haven't been able to get good results regardless