AXME
Stop a rogue agent before it sends the wrong email, deletes the wrong data, or burns your budget
Agents go rogue in ways you don't anticipate: infinite loop + tool call = $500 API bill. Wrong data scope = GDPR incident. Unexpected email send = customer escalation.
AXME Mesh provides three containment mechanisms for rogue agents: immediate halt, graceful pause, and policy-triggered quarantine.
Agents go rogue in ways you don't anticipate: 3,000 emails with a bad template, infinite loop + tool call = $500 API bill, wrong data scope = GDPR incident.
Rogue agents are predictable in hindsight
Runaway loops, wrong tool arguments, and scope creep rarely start catastrophic — they spike API volume, then errors, then spend. By the time finance notices, thousands in tokens are gone and customers may already be affected.
Containment requires targeted stop without killing healthy agents. AXME Mesh halt, pause, and quarantine act on agent identity across regions — not kubectl delete on a shared deployment.
Example: email campaign gone wrong
A campaign agent sends personalized email in batches of 100 across three Cloud Run instances. A broken unsubscribe link means 3,000 CAN-SPAM violations before anyone can Ctrl+C — there is no terminal in production.
Mesh kill blocks the agent at the gateway in under a second, preserves checkpoint state for resume, and logs who halted what for the post-incident. Policy can auto-halt on send-rate or spend thresholds next time.
SOLUTION
How teams solve this with AXME.
Runaway API spend
Loops burning tokens.
Wrong data scope
Access outside policy.
Bad outbound action
Email or delete mistakes.
PATTERNS
Production details.
Layer prevention and response
Prevention: policy-enforcement on tool allow lists, data scope, and budget caps. Detection: fleet visibility baselines and anomaly signals. Response: kill switch with audited actor and mode (halt vs pause vs quarantine).
Post-incident: export audit trail for the agent's intents and update policies — pair with /use-cases/agent-anomaly-detection/ for early warning.
Common questions
- Does halt lose in-flight work?
- Halt stops new work; paused intents can resume from last durable state after fix — mode depends on incident severity.
- Can policies auto-trigger kill?
- Yes — wire policy violations to halt or alert paths in Mesh configuration.
- How is this different from rate limiting?
- Rate limits throttle; kill switch stops the agent entirely when behavior is unsafe, not merely fast.
Related capabilities
Related links
Ship your first durable agent — in under 10 minutes.
Free tier. No credit card. Self-host or hosted — your choice.