Foundations
MDX
Levels of Agent Autonomy
A practical spectrum from suggestions to supervised actions and bounded autonomous execution.
Editorial review: clarity, operational relevance, safety boundaries, and source quality.
Assistive
At the lowest level, the system recommends an action or prepares a draft while a person performs the final step. This is often the fastest route to useful adoption because mistakes remain easy to catch.
Assistive systems are not merely prototypes. In domains with subjective quality or high consequences, draft-and-review may be the right permanent design. Measure time saved, correction rate, and whether reviewers understand the evidence; an approval click without useful context is not meaningful oversight.
Supervised
A supervised agent can perform reversible work but pauses before sensitive actions. Approval boundaries can depend on the tool, amount, data class, or confidence of the result.
Supervision works best at natural commitment points. Let the agent gather information and simulate consequences, then pause with a specific proposal. Risk-based rules can require approval for new recipients, high amounts, sensitive data, policy exceptions, or unusually low evidence quality.
Bounded autonomy
A more autonomous agent operates within explicit permissions, budgets, and monitoring. Autonomy should be earned per task through evidence, not granted to an entire product as a vague feature.
Bounded autonomy is a collection of limits: allowed resources, action types, monetary or token budgets, execution window, and stop conditions. Grant it per task and environment. A support agent proven safe at tagging tickets has not thereby earned permission to modify billing records.
Practical example
Graduating an email operations agent
The first release drafts replies only. After evaluation shows low correction rates, it may send responses for a narrow set of reversible status questions to existing contacts. Messages involving contracts, complaints, attachments, or new recipients still require approval. Expansion is tied to measured scenarios and can be rolled back independently for each capability.
Field checklist
Apply it in practice
- Classify actions by consequence and reversibility.
- Place review at the point of commitment.
- Grant autonomy per capability, not per product.
- Use production evidence to expand or revoke authority.
Decision framework
Questions to answer before you build
Autonomy is not a product-wide setting. It is a capability-by-capability decision based on consequence, reversibility, evidence quality, and demonstrated reliability.
Is assistive operation enough?
Draft-and-review may deliver most of the time savings while retaining human judgment for subjective or high-consequence work.
Where is the commitment point?
Let the agent research and prepare freely, then require approval immediately before a public, financial, destructive, or difficult-to-reverse action.
What evidence earns expansion?
Use scenario-level evaluation, correction rates, incidents, and production drift to grant or revoke a specific permission.
Common failure signals
Watch for these warning signs
- Equating fewer approval clicks with a more mature system.
- Granting write access because read-only performance looked strong.
- Using one confidence score to govern actions with very different consequences.
Selected primary references
Continue with the source material
These sources inform the wider editorial perspective for this topic. They are not presented as line-by-line citations for every statement.