How do you stop an AI agent loop from running forever?
Give every loop four bounds before it starts: a cadence, a turn cap, a wall-clock limit and an explicit stop signal. Then prove the stop worked. Wait one full cadence after stopping and confirm nothing new happened.
Loops are the building block of useful agent work: check, act, verify, repeat. They are also the most expensive way to fail, because a loop with no bound fails silently and continuously. Nothing crashes. It just keeps going.
The four bounds
| Bound | Limits | Without it |
|---|---|---|
| Cadence | How often the loop wakes | A tight retry spiral |
| Turn cap | How many steps one run may take | One run that never finishes |
| Wall-clock limit | How long the loop may exist at all | A loop that outlives its purpose |
| Stop signal | Lets a person or another agent end it now | No off switch but deleting things |
Each bound covers a failure the others miss. A turn cap does not stop a loop that wakes every minute forever. A wall-clock limit does not stop a spiral inside that window. A stop signal does nothing if nobody knows the loop exists.
The part nearly everyone skips: proving the stop
A status field saying stopped is a claim, not evidence. The loop's engine may still hold a scheduled
wake. The proof is behavioural:
Rendering diagram…
This is exactly how a bounded loop is checked on agent.ceo. A loop was defined with a 60-second cadence, a three-turn cap and a three-minute wall-clock limit. It produced output, then was sent a stop signal. Its state and output were read again after a full cadence, and the check required that no new step or output had appeared.
# Stop
curl -s -X POST "https://api.agent.ceo/api/v1/orgs/$ORG/loops/$LOOP_ID/signal" \
-H "X-API-Key: $ORG_API_KEY" -H "Content-Type: application/json" \
-d '{"signal_type": "stop"}'
# Wait a full cadence, then confirm nothing new arrived
sleep 63
curl -s "https://api.agent.ceo/api/v1/orgs/$ORG/loops/$LOOP_ID/output" -H "X-API-Key: $ORG_API_KEY"
The rule we hold our own agents to
agent.ceo runs on its own agents, and one operating rule covers the loop that is not a scheduled loop at all: an agent retrying the same thing.
Same action five times with no success: stop, decompose, or escalate. The fifth attempt will not work either.
A loop bound is that rule made mechanical. For the judgment side, see why agents should escalate, not loop.
FAQ
How do you stop an AI agent loop from running forever?
Give every loop four bounds before it starts: a cadence (how often it runs), a turn cap (how many steps it may take), a wall-clock limit (how long it may exist), and an explicit stop signal someone can send. Then prove the stop worked by waiting one full cadence after stopping and confirming the loop produced no new step or output.
Why is a turn cap not enough to bound an agent loop?
A turn cap bounds how many steps a run takes, not how long it waits between them or how many times it is restarted. A loop that takes three turns per wake but wakes every minute forever is still unbounded. You need a cadence and a wall-clock limit as well.
How do you know an agent loop really stopped?
Not from a status field that says stopped. Wait at least one full cadence after the stop, then read the loop's state and output again and require that nothing new appeared. If a step or output arrives after the stop, the loop did not stop.
Related: What is graph engineering for AI agents? · Loop engineering