AI Agent OS
Versioning AI Agent Policies
Treat permission rules and approval requirements as versioned policy so teams can explain exactly what an agent was allowed to do at a point in time.
In this guide
Practical outcomes
- Version policy changes
- Require review
- Maintain an audit history
How this works in practice
Treat permission rules and approval requirements as versioned policy so teams can explain exactly what an agent was allowed to do at a point in time.
A controlled agent separates the controller from the runtime. The controller owns the agent identity and defines policy, while a smart wallet, permission rules, quotas, relayer and scheduler constrain what the runtime can actually execute.
Implementation sequence
Turn the topic into a controlled implementation rather than a one-off transaction. Each step below should leave evidence a teammate, user or auditor can independently review.
- 01. Version policy changes. Define the expected result, capture the relevant onchain or operational evidence, and stop for review if the result differs from the plan.
- 02. Require review. Define the expected result, capture the relevant onchain or operational evidence, and stop for review if the result differs from the plan.
- 03. Maintain an audit history. Define the expected result, capture the relevant onchain or operational evidence, and stop for review if the result differs from the plan.
Evidence to retain
Record the .rama identity, controller, wallet, policy version, approval event, task identifier and resulting transaction hash. This makes a decision traceable without giving the agent unrestricted custody.
Control point
Start with small allocations, explicit recipient and contract allowlists, bounded session keys and a tested pause path. An automated workflow must never be the only place where authority or recovery knowledge exists.
Related guides
What Is Ramestta AI Agent OS?
An introduction to controlled onchain agents using .rama identity, smart wallets, scoped permissions, sponsored gas and scheduling.
Onchain AI Agents: A Practical Architecture
Learn how an AI runtime can coordinate with verifiable onchain identity, wallet permissions and bounded execution.
.rama Identity for AI Agents
Use readable .rama names to make agent ownership, discovery and wallet relationships understandable across the ecosystem.