Build with bleib.

Enforce authority outside the model

An agent should be able to reason freely while access and actions remain bounded by verifiable permission. bleib’s proposed architecture separates the two.

1. Agent request
2. Wallet permission
3. Source verification
4. Scoped result

The source or action gateway is the enforcement boundary. A model’s instructions are not a substitute for a permission check.

The integration model

• Connect a compatible agent to a user’s wallet reference. Never give it the user’s root keys.
• Resolve the goal to an approved task definition or ask the user to authorize a defined scope.
• Request a capability bound to the agent, recipient, source or action, limits, and expiry.
• Verify that capability at the source or action gateway before release or execution.
• Return the smallest approved result, and retain a verifiable record of the authority used.

Expected policy outcomes are allow, approval required, or refuse. Expired, revoked, or invalid authority must not proceed.

A small request. A narrow answer.

For a dinner plan, a source adapter could derive availability locally and release a free/busy answer. It should not return every calendar event.

{
  "agent": "marek-social-agent",
  "task": "approved-dinner-plan",
  "source": "calendar",
  "operation": "availability",
  "window": "Friday evening",
  "disclose": ["free_busy"],
  "expires": "end-of-plan"
}

A wallet the user can keep

The current proposal combines user-held root keys, portable encrypted off-chain state, and a minimal blockchain registry for independently verifiable control and revocation.

Private context and source records do not belong on a public ledger. Agent capabilities should remain narrower than the user’s root authority. Recovery and key rotation must preserve that separation.

The chain and protocol are under evaluation. No network deployment, SDK, contract address, or production API is available yet.

Where contributions can matter

• Agent interfaces and source adapters.
• Deterministic policy evaluation and verification.
• Capability scopes, delegation, expiry, and revocation.
• Portable encrypted context and compatible clients.
• Recovery, key rotation, and interoperability test vectors.
• Clear consent and permission experiences.

Start with the architecture

The architecture overview explains the separation between intelligence and authority. Nudge is the first application we’re using to test it in everyday life.