You have a kill switch.Nobody has ever pulled it.
Every stop control is a claim about how a system behaves in a situation nobody has been in. This grades what each one has actually been shown to do — and an exercise against an idle agent has shown nothing about work already running.
The stop was written by somebody imagining the incident.
It is in the runbook. It has an owner, a name and a paragraph explaining what it halts. What it does not have is a single occasion on which anybody used it against an agent that was part way through doing something. The first time that happens will be the night you need it to work, and you will find out then whether the signal reaches a process that has stopped reading its inbox.
When it was last used, what state the agent was in, what it claims to halt, whether it travels around the agent, and who can invoke it.
Two controls in the worked example are identical except for whether the agent was mid-task when it was exercised. That is the whole difference.
Including a perfect network route cut, on call, out of band, claiming to halt everything — and never once used.
Change what the agent was doing and watch a claim collapse.
This is what it has been shown to do.
9 controls, of which 6 claim more than they have been shown to do and 3 have never been used at all.
This grades a control, never a person, and says nothing about whether stopping an agent was the right call.
The full output on the worked example, verbatim.
Nine controls across five agents, one command, no dependencies. This is the complete run, pasted as printed.
STOP-AUTHORITY ASSAY - what has this control been shown to do spec_version 1.0.0 | 9 stop controls read across 5 agents CTRL CONTROL AGENT CLAIMS SHOWN DECIDED BY -------------------------------------------------------------------------------------------------------- S-01 Revoke agent credential refund_agent STOPS STOPS claimed_scope S-02 Disable scheduler entry refund_agent STOPS SLOWS exercise_conditions S-03 Console stop button support_agent STOPS SLOWS in_band_channel S-04 Feature flag kill support_agent STOPS NAMED ONLY last_exercised S-05 Vendor emergency pause outreach_agent STOPS SLOWS vendor_gated S-06 Queue drain switch outreach_agent SLOWS SLOWS claimed_scope S-07 Break-glass process kill ledger_agent STOPS SLOWS in_band_channel S-08 Runbook shutdown steps ledger_agent NAMED ONLY NAMED ONLY last_exercised S-09 Network route cut pricing_agent STOPS NAMED ONLY last_exercised CLAIMS MORE THAN IT HAS SHOWN S-02 claims STOPS, has shown SLOWS - exercise_conditions S-03 claims STOPS, has shown SLOWS - in_band_channel S-04 claims STOPS, has shown NAMED ONLY - last_exercised S-05 claims STOPS, has shown SLOWS - vendor_gated S-07 claims STOPS, has shown SLOWS - in_band_channel S-09 claims STOPS, has shown NAMED ONLY - last_exercised UNESTABLISHED ANSWERS (scored as the worst they could have been) S-07: channel - a better answer would have made this STOPS PER AGENT (read by its STRONGEST control, not its weakest) AGENT CONTROLS VERDICT STRONGEST -------------------------------------------------------------- refund_agent 2 CAN BE STOPPED S-01 support_agent 2 CAN ONLY BE SLOWED S-03 outreach_agent 2 CAN ONLY BE SLOWED S-05 ledger_agent 2 CAN ONLY BE SLOWED S-07 pricing_agent 1 CANNOT BE STOPPED S-09 ESTATE 1 of 5 agents can be stopped VERDICT: SOMETHING HERE CANNOT BE STOPPED FIX FIRST: pricing_agent (CANNOT BE STOPPED) - 1 control(s), 1 never exercised REPORTED, NOT GRADED 9 stop controls in total. How many an agent has is compared against nothing. 3 control(s) have never been used at all. 6 control(s) claim more than they have been shown to do. This grades a control, never a person, and says nothing about whether stopping an agent was the right call.
S-09 is the one worth sitting with. It is out of band, invokable by whoever is on call, and claims to halt everything. It has never been used, so it has shown nothing.
Three rules, and the reason each one is not negotiable.
A control has been shown to do what it was actually asked to do, and no more. The claim is printed beside the reading so the gap is visible, and it never raises the reading. A control nobody has used has shown nothing, whatever else is true of it.
Halting an agent that was between tasks demonstrates that new work does not start. The case you built the control for is the other one. So a claim to halt in-flight work caps at what the exercise covered — and a rehearsal against a running agent counts fully.
A signal the agent has to receive and act on works on an agent behaving normally, which is not the situation you are in. A stop that travels through the thing it is stopping cannot stop a thing that has stopped listening.
An evidence check, not a resilience review.
- A read-off of what each stop control has actually been asked to do.
- A runnable engine, a workbook that reproduces it, and a demo that agrees with both.
- A per-agent reading that names the one control currently doing the work.
- An emitter: one flag and the finding comes out as a verdict envelope.
- A check on what is still running afterwards. Queued work and retries are a different question.
- A judge of whether stopping the agent was the right call.
- A resilience or chaos-engineering programme. It grades evidence you already have.
- A scanner. Every input is an answer you supply about your own controls.
Scope: this is an engineering instrument. It grades a stop control, never a person, and asserts no regulatory position. It does not tell you whether halting an agent was the right decision, and it is not legal advice.
For the stack where an agent can act faster than you can react.
- Something in your stack runs unattended and can take actions you would want to interrupt.
- Your runbook names an emergency stop and you could not say when it was last used.
- You are about to tell somebody senior that an agent can be halted, and would like that to be true.
- You want a cheap, specific list of drills worth running this quarter.
- Nothing runs without a person watching it finish.
- You want to know what is left running after the stop. Different instrument.
- You are looking for a chaos-engineering platform. This grades evidence, it does not generate it.
The three it sits between.
Before the fact: can your gate stop the action from happening. This asks the question that only matters once it is already happening.
OpenHow bad an action is if it does run, by reversibility and reach. That decides how much stop authority is worth building.
OpenWhether an agent is fit to deploy at all. Its controls include a rollback path; this asks what that path has been shown to do.
OpenDirect answers, before you buy.
The first time you pull it
should not be the night you need it.
One purchase, lifetime access, 12 months of updates. $109, once.
Scope: an engineering instrument. It grades a control, never a person, asserts no regulatory position, and is not legal advice.
Sold by RedHub AI LLC · Secured by Stripe · redhub.ai