Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Cupertino

On Multi-agent swarming

Andrew Liang · Team

Submitted to AI Incident Response Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: On Multi-agent swarming

Open in a new tab (opens in new tab)
Share

Our contributions are as follows: We investigate whether agents assigned to individual tasks can recognize their peers and use their information without explicit collaboration instructions, and how these behaviors depend on activity-trace exposure. We measure how assertions from an authoritative planner agent affect collective answers when a worker holds contradictory evidence, distinguishing worker pushback from adoption of the planner’s incorrect answer. We test whether controlled perturbations of planner messages and worker evidence can identify planner-sensitive answers and localize influential messages without access to reasoning traces.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. Most methodologically serious entry. The deterministic replay framework is the standout, 210 byte-identical identity replays so every flip is attributable to its edit. That is a genuinely useful technique. Also honest about non-results. Weaknesses: small samples everywhere (the ROC is ten episodes), one model, and the audit's discriminative power is unresolved, not demonstrated. The conclusion overreaches, 'such a planner should not exist' is not what the data shows. Scale the question set and this is a paper.

  2. Ambitious and interesting work. The two studies carried out in this project tackle a kind of important question after the recent agent swarm incidents, primarily how information and authority move between agents. I specifically liked the planner-influence study and its implementation methods. The main limitations as stated are that all the findings rest on one model and very small screened samples so the deference rate is directional rather than settled, and the shared artifact study is more of illustrative case observations than a measured effect yet. Overall a strong, careful piece of work that would benefit most from running the planner study at the n=150-200 you mention.

Cite this project

@misc{liang2026multiagent,
  title = {{On Multi-agent swarming}},
  author = {Andrew Liang},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/on-multiagent-swarming-0h4g}},
  url = {https://apartresearch.com/sprints/projects/on-multiagent-swarming-0h4g}
}

More from AI Incident Response Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026