Agentcess
Agent trust infrastructure
Agent trust infrastructure
Every agent gets a passport. Every data interaction leaves evidence.
Agentcess is a control plane for autonomous agents: portable identity, versioned policies, budgeted runs, human approvals and a live trust map. Nothing an agent does is silent or unattributed.
How it works
Seven steps from identity to evidence
The workflow is a closed loop: an agent is registered, governed, run, supervised and audited: and the result feeds back into its trust score.
- 01
Issue a passport
Every agent gets a passport: identity, purpose, owner, and the data classes it may or may not touch. The passport is the agent's portable trust record.
ac:agent:<uuid>
- 02
Bind capabilities & tools
Capabilities and WebMCP tools are bound to the agent. Access grants scope exactly which data assets it may reach, and through which connected endpoint.
agent_tools · access_grants
- 03
Enforce policies
Versioned policies gate every action. Regulated data is refused outright; conditional actions are held for a human. The agent reports its own boundaries.
policies v1.0
- 04
Run with a budget
Give the agent a task and a dollar budget. It runs on the server and keeps going if you close your phone: each step, tool call and spent cent is recorded.
agent_runs · budget
- 05
Approve in flight
Actions that need human sign-off land in Approvals. Decisions are passkey-verified and name the person who made them: no silent overrides.
approval_requests
- 06
Collect evidence
Every data interaction leaves evidence: tool calls, decisions and attestations. Trust scores update deterministically from stored components: never invented.
evidence · trust_score
- 07
Inspect the trust map
A live trust map shows every agent, data asset and policy and the relationships between them: the governance graph of your workspace, in one view.
trust_map
The closed loop
Trust is earned, scored and inspectable
An agent's trust score isn't a random number. It is a deterministic weighted sum of stored components: policy coverage, evidence completeness, approval history and boundary adherence. Every run can only raise or lower it through real, recorded events.
- Policy coverage: does every capability have a governing policy?
- Evidence completeness: does every action leave a trace?
- Approval history: were held actions confirmed or denied?
- Boundary adherence: did the agent refuse what it must refuse?
Status lifecycle
- TrustedFull autonomy within budget and policy
- VerifiedIdentity and capabilities confirmed
- ConditionalMay run, but actions need approval
- Review requiredHeld pending human review
- RestrictedLimited to low-risk data classes
- RevokedPassport withdrawn, no actions permitted
Start governing your agents
Create a workspace, issue the first passport, and run an agent against a real budget.
Agentcess · portable identity, governance and evidence for autonomous agents