Harness notes

    Multi-agent teams

    3 min read

    A multi-agent team works by named roles, visible handoffs, and one owner per stage, with a human at the point where anything is sent.123 In the cited pipeline, nothing auto-posts.1

    Overview

    Blind peer ranking hides who wrote each answer before models rank them.4 A second agent is added only at a real bottleneck.2

    Mechanism

    1. A team takes the shape of a pipeline, a fan-out, or a specialist team, and each role has tools and a defined output.53
    2. The handoff is the objective, the artifacts, the decisions, the constraints, the open questions, and the next gate, not the transcript.2
    3. A human posts, and nothing auto-posts.1
    4. A shared channel is skipped when compute is tight or progress is not measurable.6
    5. A short model list inside one family is preferred,7 and worker outputs are merged with a constrained rule instead of one generative manager.8
    6. When workers answer the same question, each ranks the others’ answers with the authors hidden, so no worker can favor its own answer or a known model.4 Editorial note: here the ranks feed a constrained merge rather than a generative chairman.
    7. On a large codebase, each component has an owner that knows its own code and dependencies. Only the owners a change touches are woken, and a failing test goes back to the owner of the part it points at, not to whoever happens to be holding the task.9
    8. When the lead is the strongest model, it chooses each worker’s kind, size, and effort at the moment it hands work off, and it returns to a worker it has already briefed instead of briefing a new one. A handoff is not forced.10

    Applications

    The pattern applies when one agent is doing research, writing, and review thinly and the handoff is getting lost.

    Limitations

    A self-graded outcome with no independent judge conflicts with the verification pattern (see Verification and stop conditions). One chairman model writing the final answer over a blind review reintroduces a generative manager,8 and widening the pool across many vendors to get more reviewers is not the preferred way to add review.7

    Worked example

    The source repository runs three stages.4

    1. First opinions: the query goes to each model separately, and every answer is collected and shown.
    2. Review: each model receives the other models’ answers with the authors anonymized, so it cannot play favorites, and ranks them for accuracy and insight.
    3. Final response: in the repository, a designated chairman model compiles the answers into one reply.

    See also

    Further reading

    References

    Footnotes

    1. https://x.com/thegreatest_sv/status/2103457238673449245 ↩ ↩2 ↩3

    2. https://x.com/i/article/2105262478535847936 ↩ ↩2 ↩3

    3. https://x.com/cyrilxbt/status/2050214010541269225 ↩ ↩2

    4. https://github.com/karpathy/llm-council ↩ ↩2 ↩3

    5. https://x.com/av1dlive/status/2055238100079808996 ↩

    6. https://arxiv.org/abs/2609.21032 ↩

    7. https://arxiv.org/abs/2609.17306 ↩ ↩2

    8. https://arxiv.org/abs/2609.09815 ↩ ↩2

    9. https://arxiv.org/abs/2610.07832 ↩

    10. https://x.com/i/article/2104968643021045760 ↩

    Graph

    Backlinks