CORE GUIDE
Bounded autonomy
Give an agent enough freedom to complete useful work, but place hard boundaries around cost, time, permissions and irreversible effects.
Mental model
Autonomy is an action budget: the agent can operate inside a defined envelope and must stop, ask, verify or escalate when it reaches a boundary.
Why it matters
Unbounded loops can spend indefinitely, repeat side effects and drift away from the user's intent. Over-constrained agents are little more than fixed workflows. The engineering goal is to match autonomy to the reversibility and impact of each action class.
01
Build the autonomy envelope
Boundaries can include maximum steps, tool allowlists, spending caps, timeouts, approval requirements and confidence-based escalation. Low-risk reversible actions may be automatic, while high-impact or ambiguous actions require stronger evidence or a human decision.
02
Example: a support agent
The agent may search the knowledge base, draft a reply and issue a small predefined credit automatically. A large refund requires approval, while conflicting policy evidence forces escalation. The same agent therefore has different autonomy levels for different actions.
Common failure modes
- Assigning one autonomy level to every tool and action.
- Allowing retries to bypass approval or spending boundaries.
- Adding vague 'human in the loop' steps without explicit escalation criteria.
Engineering heuristics
- Classify actions by reversibility and impact.
- Set finite step, time and cost budgets.
- Treat escalation as a normal success path, not an agent failure.
Takeaways
- 01More autonomy is not the same as more capability.
- 02Boundaries should scale with action risk.
- 03A reliable agent knows when not to continue.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.