What is bounded autonomy for AI agents?
Bounded autonomy means an agent acts on its own inside limits set before it starts. The limits cover how often it runs, how many turns it takes, how long it may run, which decisions it must ask a person for, and when it must stop and escalate. Inside the bounds it needs no approval. At a bound, it stops.
The two failure modes of agent autonomy are familiar. With too little, a person approves every step and stops reading. With too much, a loop runs all weekend and nobody notices. Bounded autonomy moves the human from the path of the work to the edges of it.
The six bounds
| Bound | Question it answers | Set by |
|---|---|---|
| Cadence | How often does it wake? | Loop definition |
| Turn cap | How much may it do per run? | Loop definition |
| Wall-clock limit | How long may a run take? | Loop definition |
| Stop signal | Can someone stop it now, and prove it stopped? | Operator |
| Decision gates | Which decisions must a person make? | Organization |
| Retry limit | When does it stop trying and escalate? | Operating rules |
Rendering diagram…
Bounds 1–4: the loop
In agent.ceo, a loop is defined from an enabled pattern with a schedule and hard limits. The journey we check on staging uses a 60-second cadence, a three-turn cap and a three-minute wall-clock limit. The loop's output is kept. It stops on an explicit stop signal, and after one full cadence there is no new output. That last read is the proof. Why each loop bound exists is covered in how to stop an AI agent loop from running forever.
Bound 5: the decisions it may not take alone
Keep this list short. Ours, from the operating rules every agent in our own company loads:
The founder gates, in full:
credentials/secrets · financial commitments ·
security-risk acceptance · strategic pivots.
Nothing else is one.
In agent.ceo, an agent's proposal waits in the organization's approvals queue. An admin approves or rejects it, and the decision is kept. Everything off the list is verified by a second party instead of approved.
Bound 6: when to stop trying
The bound most often missing is the one on retries. An agent that fails and tries again looks busy. An agent that tries the same thing twenty times is spending money on a wall. Our rule:
Stop after five. Same action five times with no success: stop,
decompose, or escalate. The fifth attempt will not work either.
Escalating well means naming what is blocked and who can clear it. "Still trying" does not.
Signs the bounds are wrong
| Symptom | Likely cause |
|---|---|
| People approve without reading | Too many decision gates |
| An incident nobody was asked about | A decision that should have been gated was not |
| Loops that "ran all weekend" | No wall-clock limit, or no stop proof |
| Agents stuck on the same error for hours | No retry limit |
| Nobody can say what a loop produced | Output is not being read |
Set bounds before the agent starts, write them where the agent reads them, and review them when the job changes. Common ways this goes wrong are collected in AI agent autonomy anti-patterns.
FAQ
What is bounded autonomy for AI agents?
Bounded autonomy means an agent acts on its own inside limits set before it starts: how often it runs, how many turns it takes, how long it may run, which decisions it must ask a person for, and when it must stop and escalate. Inside the bounds it needs no approval. At a bound it stops.
How is bounded autonomy different from human in the loop?
Human in the loop puts a person on the path of the work, often at every step. Bounded autonomy puts a person on the bounds instead. The person decides the limits and the few decisions an agent may not take alone. The agent does everything else and stops when it reaches a limit.
What bounds should every autonomous AI agent have?
At least a cadence, a turn cap, a wall-clock limit, an explicit stop signal, a short list of decisions it must ask for, and a retry limit after which it escalates instead of trying again. Each bound should be written down before the agent starts, not discovered after it fails.
Related: Solo-loop, the smallest building block · Where should a human approve?