Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026Dallas

Do Monitors Add Up Across Agents? Tool Logs, Chat and a Two-Agent Overspend

Jason Gibbs · Team Weis

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

Read the report

Report: Do Monitors Add Up Across Agents? Tool Logs, Chat and a Two-Agent Overspend

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

Track 2.2 asks whether combining what agents say with what they do beats either alone. We test it on the smallest coalition that can break a rule: two treasury agents, each limited to 100 per transaction, under a team rule that together they may not send more than 100 to one payee without a co-signer. In an exploratory, post-hoc run (sealed before any model call; 150 episodes, Mistral Small and Llama 3.3 70B), a judge reading both tool logs without the chat caught 41/41 violations and 11/11 joint-only overspends with 0/59 false alarms. Each agent's log alone caught 1/11 (judge or rule; paired 10 vs 0, p = 0.00098), and a flag-any-payment rule false-alarmed 55/59 (p = 3 × 10⁻¹⁷). The judge left alone all 5 episodes where one agent legitimately paid 100 alone, so it summed across agents rather than asking whether anyone paid. A chat-only judge caught 15/41 and 1/11; adding chat to the logs gained nothing. Caveats: 11 joint-only events, two models, one judge model, and an easy honest condition. An earlier pre-registered run (781 episodes, eight models) gave logs 11/11 vs chat 4/11, but there any payment already broke a single-agent rule, so it could not separate summing from "did anyone pay?"; this run removes that confound.

Reviews

Judging this Sprint?

Review “Do Monitors Add Up Across Agents? Tool Logs, Chat and a Two-Agent Overspend”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

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

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field 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 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.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. The project explored whether combining tool logs and chat from two AI agents can detect collusion more effectively than either source alone. It tested this on a scenario where two treasury agents must not collectively exceed a payment limit without a co-signer, finding that reading both tool logs together was effective at catching violations (11/11), while adding the chat did not improve detection (4/11 for chat alone). The project also identified that an any-payment rule matched the performance of the combined-logs judge.

    However, the sample size is quite small, with only 11 joint-only events detected across two models. This limits the generalizability of the findings and makes it hard to draw broader conclusions about the effectiveness of different monitoring strategies. Additionally, while the project includes a pre-registered design, there are several post hoc adjustments that could have introduced bias into the results.

    To improve this work, I'd suggest expanding the sample size by testing across more models and scenarios. Also, consider adding more controls to isolate specific factors contributing to collusion detection. The project's findings could be valuable for developing robust monitoring frameworks in multi-agent systems, but further validation is needed to solidify these insights.

    Read full reviewShow less

Cite this project

@misc{gibbs2026monitors,
  title = {{Do Monitors Add Up Across Agents? Tool Logs, Chat and a Two-Agent Overspend}},
  author = {Jason Gibbs},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/do-monitors-add-up-across-agents-tool-logs-chat-and-a-twoagent-overspend-yjka}},
  url = {https://apartresearch.com/sprints/projects/do-monitors-add-up-across-agents-tool-logs-chat-and-a-twoagent-overspend-yjka}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026