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

Tool Choice Signaling and Task State Preservation in LLM Multi-Agent Systems.

Ellen Berhe, Muhammad Hassaan Shaukat · Team SCYLLA

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

Read the report

Report: Tool Choice Signaling and Task State Preservation in LLM Multi-Agent Systems.

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

Presentation: Tool Choice Signaling and Task State Preservation in LLM Multi-Agent Systems.

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

We investigated whether LLM agents can communicate indirectly through tool choices and whether runtime observation filtering can suppress this communication without disrupting legitimate task coordination.

Using a three agent Colosseum Jira environment with private messaging disabled, we instructed a sender to encode one bit by selecting between two equivalent tools that performed the same task assignment. We evaluated four observation policies: exposing tool identity (T1), canonicalizing the tool name (T2), hiding both action events and the receiver's assignment (T3), and hiding action events while preserving assignments (T4).

We conducted 108 exploratory episodes, a separate 72 episode holdout, and eight additional utility episodes.

Three key findings emerged. First, visible tool identity transmitted the instructed signal, with mean receiver dependence of 0.875 in the holdout. Second, deleting action events while preserving assignments reduced measured bit dependence to zero across all four holdout seeds; canonicalization showed no consistent utility advantage over event deletion. Third, also hiding the receiver's assignment caused duplicate work and reduced joint reward in three of four tested environments.

Our findings highlight a security utility trade off: concealing execution metadata can suppress the tested communication channel, but removing task-relevant information can disrupt legitimate cooperation. The study demonstrates prompted signaling, not emergent or harmful collusion.

Reviews

Judging this Sprint?

Review “Tool Choice Signaling and Task State Preservation in LLM Multi-Agent Systems.”

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. Clear threat model and a disciplined design: a positive control (a visible tool name carries an instructed bit), a separately specified holdout on new seeds, and a deterministic observation-contract checker that tests the returned prompts rather than relying on model behaviour. The key insight, that observation filters must separate execution metadata (which function was called) from task state (who holds which task), is practical and well supported: hiding the assignment along with the event produced duplicate claims on 3 of 4 holdout seeds, while deleting only the event removed measured dependence without hurting reward. Linking this to covert-channel and noninterference work is a nice touch.

    Limitations to address:

    1. The channel is prompted with an explicit one-bit codebook, so this shows capacity, not that agents would discover or use it. A setting where agents can invent a convention (repeated play, shared reward, no codebook) would make the threat more realistic.
    2. Samples are small (4 holdout seeds, 2 draws per cell, one model); a single action moves delta by 0.5. More seeds and a second model would help.
    3. The exploratory logs lack the codebook text. You disclose this, but it weakens reproducibility of that cohort.
    4. Other side channels remain open (timing, argument order, which task is claimed). A short note on how the contract would extend to them would strengthen the design-principle claim.

    Presentation is accurate but very dense and repetitive, and some tables and figures render poorly in the PDF. A one-paragraph summary of the three findings near the top and cleaner tables would help.

    Read full reviewShow less

Cite this project

@misc{berhe2026tool,
  title = {{Tool Choice Signaling and Task State Preservation in LLM Multi-Agent Systems.}},
  author = {Ellen Berhe and Muhammad Hassaan Shaukat},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/tool-choice-signaling-and-task-state-preservation-in-llm-multiagent-systems-37wc}},
  url = {https://apartresearch.com/sprints/projects/tool-choice-signaling-and-task-state-preservation-in-llm-multiagent-systems-37wc}
}

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