agent · AVC

Agents participate inside the operating system.

Identity, capability, authority, evidence and runtime state remain visible.

Agent boundary

Capability is explicit. Authority is never inferred from intelligence.

The model, agent runtime and owner occupy different trust positions.

Model

May
Generate an inference within supplied context.
May not
Hold credentials, approve authority or claim execution evidence.

Agent

May
Plan and execute named capabilities under policy.
May not
Expand its grant or conceal blocked state.

Specialist

May
Perform one bounded role with a scoped capability bundle.
May not
Delegate beyond the parent mission boundary.

Owner

May
Approve, reject, narrow, pause or revoke consequential work.
May not
Transfer non-delegable ownership to the agent runtime.
Governed actor

An agent is an identified participant with bounded capabilities.

Identity, model, provider, tools, authority and runtime state remain separate facts rather than one vague notion of intelligence.

Delegation

Specialists and subagents inherit narrower scope, not wider authority.

Parallel work is reserved for independent tasks with clear ownership and evidence requirements.

Accountability

Every meaningful agent action produces events, evidence or an explicit block.

Silence is not success, and fluent output is not proof that the requested work was authorized or completed.