Who acted?
Every worker and service has a stable identity and role. Shared anonymous automation is rejected.
Governed agents
Agents may work. They do not receive silent authority.
Agentic AI is optional. When justified, each worker receives a named identity, narrow tools, a bounded task, a budget, a timeout and an accountable human gate.
What could it do?
Tool allow-lists, isolated workspaces, least privilege and explicit approval before material actions.
What happened?
Claims, handoffs, tool outcomes, decisions, timeouts and stop events form a durable operating record.
The agent contract
- Purpose
- one named workflow and success condition
- Inputs
- approved data classes and source boundary
- Tools
- allow-listed operations only
- Approval
- human gate for material or regulated decisions
- Limits
- budget, time, retries and concurrency
- Stop
- operator halt, timeout and fail-closed behaviour
- Evidence
- receipts, outputs, negative findings and owner
- Recovery
- rollback or continuation from durable state
Synapse Channel is relevant public evidence for identities, claims, mailboxes, receipts and local coordination. Experimental surfaces remain outside delivery guarantees unless separately validated.
Where agents should stop
- autonomous legal, medical, credit or investment decisions
- unreviewed external publication or financial execution
- credentials or tools broader than the named task
- silent model or policy changes
- work without a receiving owner, timeout or recovery path