Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 12, 2026Portland, ME

Safety always wins: a harness-layer emergency stop for agent meshes

E SMITH · Team TYMBAL_AI

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

Read the report

Report: Safety always wins: a harness-layer emergency stop for agent meshes

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

Safety always wins: a harness-layer emergency stop for agent meshes - Hit the "E STOP" but - making SAFETY FIRST and easy route.

agent-estop is a harness-layer control that (1) halts the feed of work to every agent in a mesh with one command, (2) as a marker on durable storage that survives a harness restart or host reboot, (3) queues work issued during the halt and drains it only when a human resumes, (4) under a three-level hierarchy in which a lower level never clears a higher one, so a scheduled unblock cannot undo a safety estop, and (5) detects nothing: it is the stop other controls call. Where the harness exposes an interrupt channel it also cancels the running turn; halts can be scoped by agent name. Against the nine documented phases of the July 2026 Hugging Face intrusion it interrupts three directly and one conditionally; five ran from footholds outside any harness. Generalized from a control in production since March 2026; ~700 lines of shell, tested under bash and zsh.

Reviews

Judging this Sprint?

Review “Safety always wins: a harness-layer emergency stop for agent meshes”

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 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.
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 report is honest about its scope and limitations, and Appendix A is especially clear. Its strongest contribution is the ownership-based safety model, while the channel-drift example and vendor requirements are practical and useful. The report also clearly separates prior work from new contributions. However, the approach depends on another control detecting the threat first, so the authors should show exactly when it would activate during an incident replay. Measuring coverage by action count may also hide the seriousness of the uncovered phases. The phase totals add up to 16,521 rather than approximately 17,600, and Phase P-3 appears only partially covered. A feed halt may not stop actions already running during a long autonomous turn. The report should compare this approach with infrastructure-level controls and improve independent verification, since much of the current evidence is self-reported. The abstract, internal identifiers, control-level definitions, duplicated tables, and unclear terminology also need revision.

    Read full reviewShow less

  2. I think this project addresses a practical problem because stopping a group of agents should not depend on those agents choosing to cooperate. It places the emergency stop in the software that assigns their work and records which agents should remain stopped. The paper clearly distinguishes this from stopping processes already running or reversing actions on external systems.

    The central question is whether that stop remains effective when other work resumes. The current implementation can let queued tasks proceed even while the affected agents are supposed to remain stopped, so I would not yet treat it as a dependable emergency stop. The paper offers a small prototype with clearly stated limits, but its main protection needs stronger support before it can be reliable.

Cite this project

@misc{smith2026safety,
  title = {{Safety always wins: a harness-layer emergency stop for agent meshes}},
  author = {E SMITH},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/safety-always-wins-a-harnesslayer-emergency-stop-for-agent-meshes-o9me}},
  url = {https://apartresearch.com/sprints/projects/safety-always-wins-a-harnesslayer-emergency-stop-for-agent-meshes-o9me}
}

More from AI Incident Response Sprint

All 231 entriesBrowse all projects

Build something like this at the next Sprint

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