§ PLATFORM / KILL SWITCH
Stop one agent. Keep the proof.
When an AI agent misbehaves, you need it stopped before its next action, not after the next deploy. Execlave pauses a single agent from the dashboard, CLI or API. The pause takes effect server-side in single-digit milliseconds, the agent's next policy check is blocked, and the audit trail records who stopped it, why, and every call it tried afterwards.
§ CAPABILITIES
A kill switch built for evidence, not just for emergencies
Stopping an agent is the easy half. Showing an auditor that it stayed stopped is the half most teams cannot do.
A kill switch built for evidence, not just for emergencies
01 / Capability
One agent, one action
Pause a single agent from the dashboard, with the CLI command execlave agents pause or through the REST API. Other agents keep running. The reason you give is stored with the pause.
02 / Capability
Synchronous and fail-safe
Every enforce call checks the agent's status first. Only an agent that is explicitly active may run: paused, archived or unknown all block. The first check after the pause commits is already denied.
03 / Capability
Evidence in both directions
Pausing and resuming write hash-chained audit records with who, when, the reason and the status before and after. Each call the paused agent attempts is recorded as a denial with the reason kill_switch.
04 / Capability
Wired into your stack
Open dashboards update live the moment an agent is paused. Agents with a monthly budget can pause themselves when it runs out. Resuming is reserved for admins and owners, so a stop cannot be quietly undone.
§ MEASURED
How fast the pause takes effect
The server-side cost of the pause: the real pause transaction, database included, under the same row-level-security role production uses. A client calling over HTTP also pays authentication and network round-trip time.
How fast the pause takes effect
§ QUESTIONS
AI agent kill switches, answered
AI agent kill switches, answered
What is an AI agent kill switch?
An AI agent kill switch is a control that stops a running agent from taking further actions, immediately and without redeploying it. In Execlave it is a pause on one agent: the agent's status changes to paused, and every policy check it makes afterwards is denied until an admin resumes it.
How fast does Execlave's kill switch take effect?
Server-side, the pause transaction takes 2.9 ms at p50 and 5.4 ms at p99, database included. Propagation is synchronous: the first policy check after the pause commits is already blocked. A client calling over HTTP additionally pays authentication and network latency, which depends on its distance to the API.
Does the kill switch stop an agent that is not instrumented?
No. The pause is enforced at the policy check, so it only stops agents that call Execlave before they act, through the SDK or the enforce API. An agent that never checks in is outside the control.
What happens if the agent cannot reach Execlave?
By default the SDKs fail open: if the governance service is unreachable, the action proceeds without a server decision, so a pause cannot hold during an outage. Set enforcementOnOutage to fail_closed (enforcement_on_outage in Python) for agents that must stop whenever Execlave cannot confirm they are allowed. The JavaScript SDK also reports each bypass window to your audit trail.
Who can pause and resume an agent?
Pausing needs pause permission on that agent. Resuming needs an organization admin or owner. The asymmetry is deliberate: stopping an agent should be easy, restarting it should be a decision. Both actions are written to the audit trail with the user and the reason.
Does a kill switch help with the EU AI Act?
Article 14 expects human oversight of high-risk AI systems to include a way to interrupt the system, a stop button or similar procedure. Execlave's pause is such a procedure for agents, and its audit records are evidence that it exists and was used. It supports your oversight measures; it does not make a system compliant on its own.
§ COMPARISON
Execlave's kill switch vs. the usual ways to stop an agent
Revoking a key or rolling back a deploy does stop an agent. Neither tells you why it was stopped or proves that it stayed stopped.
Execlave's kill switch vs. the usual ways to stop an agent
| Aspect | Revoke the API key | Redeploy or feature flag | Execlave kill switch |
|---|---|---|---|
| Scope | Every service that shares the key | The whole deployment | One agent |
| Takes effect | Next authenticated request | When the rollout finishes | Next policy check, synchronously |
| Reason recorded | Only if someone writes it down | In a commit or ticket, if at all | Stored with the pause and in the audit trail |
| Proof it held | Rejected requests in server logs, if kept | None | Every blocked call audited |
| Restart | Issue and distribute a new key | Another deploy | Admin resume, audited |
Give every agent an off switch before you need it
Register an agent, add one enforce call, and the kill switch works from the first request. Cloud or self-hosted.