Stop Micro-Managing AI Agents: Claude Code's Opus 5.5 Leads, Haiku 5.5 Works in Parallel, Zero Regressions

Anthropic put something new into Claude Code: Agent team mode. One Opus 5.5 acts as lead, three Haiku 5.5s act as teammates, tasks are broken into a DAG, teammates claim tasks themselves, and they communicate over a peer p2p bus. Opus only signs off at the end. Test results: all 12 tasks were completed, the security audit found a timing leak in JWT claims, the performance optimization of the Redis lock made it 2.8x faster, and QA ran the full test matrix with 58/58 green and zero regressions.

The launch command is just one line:

claude --model opus --teammate-mode auto

The thinking behind this setup aligns with what mirku discusses in “Harness Engineering.” mirku argues that when model output goes wrong, the problem is often not in the model weights but in the execution environment. The environment is the harness. When Dario Amodei launched Claude Code, he said you need an interface and you need a harness to use models. Models are reasoning engines, but they cannot build a stable operating system by themselves. The harness controls what the model reads, which tools it can trigger, which data survives context resets, and when it must stop.

A model in an empty chat box will only generate small talk. Put it in a repo with a terminal, test suite, project map, and isolated workspace, and it can deliver usable software. The weights did not change; the harness did.

The full tree structure Grirku gave is as follows:

> Opus 5.5 team lead builds the task DAG and coordinates
> Haiku 5.5 x3 teammates - security / perf-scout / qa-engine
> peer p2p bus for direct mailbox and self-claim tasks
> atomic mutex DAG locks files with zero leader polling
> JEV fork layer routes deterministic decisions
> Opus signs off only after consensus and zero regressions

Paste this tree and the prompt below into Claude Code:

Rebuild my Claude Code setup around this tree:

Run the main session on Opus 5.5 with effort xhigh and --teammate-mode auto.
Spawn three Haiku 5.5 teammates: security, perf-scout, and qa-engine.
Let them self-claim tasks from a shared DAG and communicate over a peer p2p bus.
Route deterministic forks (which file, which model, retry or stop) through JEV.
Use an atomic mutex DAG so teammates never collide on the same file.
Opus only coordinates the plan and signs off at the end. Teammates handle the work.
Show me every change as a diff first. No edits until I say go.

Several design choices here are worth noting:

• peer p2p bus: teammates communicate directly with each other without going through the lead. The lead does not need to poll and will not become a bottleneck.
• atomic mutex DAG: file-level locking, so multiple teammates writing to the same file never collide.
• JEV fork layer: takes deterministic decisions such as “which file, which model, retry or stop” out of model weights and routes them through lower-level mechanisms.
• Opus only coordinates the plan and signs off at the end. The concrete work is handled by the Haikus.

One commenter under the tweet wrote: “So now opening a few more windows isn’t what makes it multi-agent.”

mirku breaks reliable AI systems into three layers: the harness layer, the loop layer, and the graph layer. This team mode maps exactly to them:

• Graph layer: Opus builds the task DAG, Haikus claim tasks themselves, and long tasks are split into isolated contexts for execution.
• Loop layer: QA-engine runs the test matrix, perf-scout runs performance analysis, and results are validated with mechanical evidence rather than relying on the model saying it did it right.
• Harness layer: atomic mutex locks files, JEV routes decisions, and peer p2p bus manages communication—all outside model weights.

If you have also run into the situation where “the model keeps forgetting the rules and changing the prompt doesn’t help,” the problem may not be the prompt. The problem may be the harness. Instead of continuing to add adjectives to the system prompt, check the operating system around the model.

The main tweet’s last line is “save it, it’ll come in handy.” — save it; it will be useful.

The video is the terminal interface from an actual run, where you can see the Opus lead, three Haiku teammates, and DAG task status:

Stop micro-managing AI Agents. Set up the task DAG, wire in the verification mechanisms, and let the models do the work themselves. You only need to review the results at the end.