Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026

Reading Between the Agents: Message and Action Evidence in Coordination Monitoring

Mahlet Molla, Lydia Lee · Team Decode

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

Read the report

Report: Reading Between the Agents: Message and Action Evidence in Coordination Monitoring

Open in a new tab (opens in new tab)
Code (opens in new tab)More on github.com (opens in new tab)
Share

Monitoring multi-agent coordination requires weighing what agents say against what they do, not conflating evidence with intent. In a pilot study of a two-option game where the operator prefers option A, we first varied scripted sender messages and measured two language models' responses as receivers. Sonnet 4.6 chose the permitted but discouraged option B in 8/20 trials under a B-directed proposal, GPT-5.4-mini in 7/20 under an A-directed one; however, direction was bundled with wording. We then compared four monitors: message-only, action-only, combined, and combined with a checklist. Message-only and combined monitors reached AUROC 1.00 on the resolved subset, so combining evidence showed no gain. This reflects a ceiling, not equivalence, because the endpoint is post hoc and perfectly separable by template. These findings show how endpoint construction, evidence asymmetry, and prompt design can limit what a monitoring evaluation shows, and we offer two diagnostics, saturation and evidence-contract checks, for future benchmarks.

Reviews

Judging this Sprint?

Review “Reading Between the Agents: Message and Action Evidence in Coordination Monitoring”

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. Good question (does adding action evidence help a coordination monitor beyond messages?) with useful related-work framing. The write-up is honest about why the monitor comparison could not answer it: endpoint E is perfectly separable by message template, so message-only monitors saturate at AUROC 1.00 and there is no headroom for action evidence to help; and the M2 action-only monitor was asked about messages it could not see, so its all-zero scores are not an action-only result. Turning these into explicit lessons (the saturation and evidence-contract checks) is a good move.

    To make the next version informative:

    1. Define the endpoint from something not recoverable from the message alone, for example whether the pair actually ended at B/B after a B-directed message, or independent labels on live (non-scripted) sender conversations.
    2. Include hard cases where messages and actions disagree: a B proposal that neither agent follows, and B/B outcomes with no proposal. These force message-only and action-only monitors to diverge.
    3. Write a separate monitor prompt for each evidence view so each is asked only about what it can see.
    4. Use minimal-pair templates that differ only in the target option, so target is not confounded with wording, length and payoff mentions.

    The receiver result is intriguing (Sonnet chose B less often after a B-directed proposal; GPT-5.4-mini chose B less after an A-directed one) and might deserve a short follow-up with matched wording. With 20 repeated draws of one fixed game, present it as anecdotal for now.

    Read full reviewShow less

Cite this project

@misc{molla2026reading,
  title = {{Reading Between the Agents: Message and Action Evidence in Coordination Monitoring}},
  author = {Mahlet Molla and Lydia Lee},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/reading-between-the-agents-message-and-action-evidence-in-coordination-monitoring-9dpn}},
  url = {https://apartresearch.com/sprints/projects/reading-between-the-agents-message-and-action-evidence-in-coordination-monitoring-9dpn}
}

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