Glossary
What is a harness autonomy tier?
A harness autonomy tier describes the level of independence an AI agent has inside the harness that orchestrates it, from every action human-approved to agents acting fully on their own authority inside policy you wrote.
§ 01 Definition
Same record at every tier.
A harness autonomy tier describes the level of independence an AI agent has inside the harness that orchestrates it. The tiers form a ladder, from every action human-approved to agents acting fully on their own authority inside policy you wrote. The harness coordinates the work at any tier; what changes between tiers is how much the human sits in the loop.
The substrate produces the same attributable, tamper-evident record at every tier. A supervised agent and a sovereign agent write to the same audit trail; the difference is in the policy that governs them, not in the record they leave behind. Trust is earned by watching the trail accumulate at the lower tiers and promoting workflows when the evidence supports it.
§ 02 Questions
Harness autonomy tiers, answered.
What are the standard autonomy tiers?
A common four-tier ladder: Supervised (every action proposed; a human approves), Coordinated (agents hand off to each other within policy, every handoff attributed), Autonomous-within-Policy (agents decide and act inside bounded policy; humans on exceptions), and Air-gapped (sovereign deployment with no external dependency). See the agent classes for how Subnet345 frames this.
Why do the tiers matter for compliance?
Different regulatory frames expect different default tiers. High-risk classifications under the EU AI Act expect documented human oversight, which presses toward Supervised or Coordinated at first. NERC CIP on Bulk Electric System assets presses toward lower-autonomy tiers for any consequential action. The substrate's audit trail is what lets an operator defend the choice of tier for a specific workflow.
Can the substrate's audit trail support promotion between tiers?
Yes. Promotion is a decision an operator can defend with the accumulated trail at the lower tier: which actions the agent took, what the human approved, what the policy refused, what was escalated, how the outcomes compared to the prior process. The substrate makes promotion a defensible decision rather than a leap of faith.
§ 03 Related
Where the term lives.
Regulatory frames
Related