Back to guides

Guide · Agents

Agent Autonomy Limits and Kill Switches

Unbounded agents loop, spend, and escalate. Production agents need charters, autonomy caps, and a real kill switch—not hope.

Unbounded agents fail Agent Governance—charters, budgets, and kill switches are mandatory in production. Primary control: Agent Governance

Charter before code

Write one page per agent class before it touches production tools:

- Purpose — what "done" means

- Allowed tools — names, not "whatever MCP is configured"

- Data classes — what it may read/write

- Forbidden actions — e.g. prod deploy, payment capture, mass email

- Budgets — max steps, wall time, token/spend ceiling

If you cannot fill that page, you are not ready for Agent Governance—you have a demo with credentials.

Runtime bounds beat prompt manners

Caps belong in the orchestrator: step counter, timeout, spend meter, checked every turn. "Please don't loop" in the system prompt is not a control. Pair with Tool Safety allowlists and separate identities per agent class so a summarizer never shares an admin token.

A coding agent with shell + `git push` in CI, no max steps, and a README injection once opened a storm of PRs. The only interrupt was revoking the GitHub app. That is the failure mode you design against.

Kill switch drill

A kill switch is a control-plane flag (or queue drain) that:

1. Stops scheduling new steps immediately

2. Cancels in-flight tool calls where the platform allows

3. Is reachable by on-call without a redeploy

Test it in staging the way you test failover. Document the runaway-agent runbook next to infra SEVs. Ambiguous or high-impact plans should already escalate to Human Approval—kill switch is for when autonomy was wrong, not for every button click.

Assess against APRF Core

Run the Core Profile quiz — gated pass/fail blockers for AI production readiness, not a vanity score.

Frequently asked questions

Is a kill switch just "turn off the feature flag"?
That is the minimum. Prefer a flag checked every agent step plus cancellation of in-flight tool calls where possible.
How many steps should an agent get?
Start low for production (tens, not thousands). Raise only with eval evidence and monitoring.
Do chatbots need this?
Single-turn chat needs less. Anything with tools, loops, or A2A needs Agent Governance controls.