Register & delegate
Establish the principal, agent identity and approved mandate.
Agent Passport · development & pilot programme
Make authority
specific to the action.
Agent Passport is Legis Labs’ approach to connecting an AI agent with an accountable principal, a bounded mandate and evidence of the decisions made at execution.
What the passport
is designed to carry.
A credential supports a decision. The receiving system must still validate it and enforce its own policy.
Accountable principal
The person or organisation granting authority, with the appropriate identity or business checks.
Agent identity and key
A verifiable binding to the agent or workload, including issuer and credential status.
Bounded mandate
Permitted tools and actions, resources, financial limits, expiry and restrictions on delegation.
Approval and revocation rules
When a new human approval is needed and how a withdrawn mandate takes effect.
Evidence references
References to supporting checks and policy versions, with sensitive personal data kept out of portable credentials.
Interactive concept
Change the action.
See the decision.
This browser-only illustration uses synthetic data and a simplified policy. It does not verify an identity, sign a credential, connect to a bank or move money.
A supplier payment request
Agent payments-assistant-07 · Example Treasury Ltd
Illustrative policy: up to £1,000 may proceed; £1,001–£5,000 needs a new human approval; above £5,000 is outside the mandate. Inactive mandates and unapproved suppliers are blocked.
POLICY DECISION / DEMO V1
Allow within scope
The mandate is active, the supplier is approved and £750 is within the £1,000 autonomous limit.
- Agent
- payments-assistant-07
- Mandate
- supplier-payments-example
- Policy version
- demo-v1
- Execution
- None · illustration only
- Credential signature
- Not implemented in this demo
The downloaded JSON is an unsigned example, not audit evidence from a deployed system.
Proposed pilot architecture
Put the check
where it can stop the action.
The pilot should connect to an authorised tool or API boundary. A check in an agent’s prompt alone is insufficient to enforce permissions.
Validate & decide
Check the actual request against current authority and policy.
Enforce & record
Allow, deny or hold for approval, then record the observed result.
Designed for existing infrastructure
Our proposed approach is intended to build on established identity and payment patterns. Interoperability targets include OAuth-based delegation, MCP tool access, agent discovery and payment mandates. Compatibility must be demonstrated for each integration.
Measured in a bounded workflow
Start with one workflow, named tools and agreed test cases. Measure policy coverage, bypass attempts, revocation, false blocks and time to reconstruct a decision before expanding scope.
Questions we expect
Clarity before deployment.
Does an Agent Passport make an AI system safe?
A passport can establish identity and delegated scope. Safety also depends on enforcement, accurate evidence, secure tools and the wider workflow. It cannot guarantee correct reasoning or prevent every harmful outcome.
Does this eliminate hallucinations?
No. Explicit policy checks can reject defined invalid actions, while evidence checks can identify unsupported claims within their scope. An LLM can still produce inaccurate output, including output that passes an incomplete policy.
Is the Agent Passport available in production?
This site presents a development and pilot programme. Production capabilities, integrations, service levels and independent test results must be agreed and evidenced for a specific engagement.
Does a passport replace the bank’s controls?
No. The institution retains its policy, approval obligations and control of execution. Agent access requires the institution’s permission and supported integration.
Work with Legis Labs
Put a real workflow to the test.
Bring an agent use case, an accountable owner and a decision you need evidence to support.