03 / GOVERNANCE

The model proposes. Authority lives elsewhere.

AI reasoning can be flexible. Enterprise execution must remain controlled. Identity, permissions and approval come from the trusted execution layer — not from what an LLM says.

Explore the boundary

Staging authority does not include production.

The same proposed action can be useful and unauthorized. A policy decision must happen before execution, not after it.

A request is not permission.

Interactive illustration only. This runs in your browser and is not connected to any infrastructure. Try the request, then approve that exact revision as the simulated human reviewer.

Trusted identity
workflow-agent-17
Authorized scope
staging/*

DENIED · production/payment-api is outside staging/*. No human approval is recorded.

Illustrative evidence: identity workflow-agent-17; target and revision; policy decision; approval binding. After a real action, verify observed state against approved intent and reconcile any mismatch.

A restart of production/payment-api by workflow-agent-17, authorized for staging/*, is denied. Explicit human approval for that target, restart operation and revision 182 permits only that exact request. Identity and permissions come from the trusted execution layer, never from what an LLM claims.

What the boundary holds

Check the request. Preserve the evidence.

Define the controls for each workflow as part of the integration, including how to stop and who can approve an exception.

  1. Identity & authorization

    Use trusted workload identity and explicit permissions. A model cannot inherit another agent's authority by claiming its name.

  2. Scope & approval

    Bind tool access and changes to approved targets, operations and revision. Changed intent requires fresh approval.

  3. State & audit

    Record the request, policy decision, approval and execution outcome so a reviewer can reconstruct what was authorized and what happened.

  4. Verify & reconcile

    Compare observed state with approved intent. Surface partial failures, unexpected changes and unresolved outcomes for accountable follow-up.

Exception-driven review

Human attention goes where it matters.

AI handles volume. Humans handle judgment. Review conditions are explicit, not a request to blindly trust a confidence score.

  • Normal result

    Process within the pre-approved policy.

  • High-value transaction

    Require human review.

  • Missing evidence / conflicting data

    Pause and ask a responsible reviewer.

  • Low confidence

    Escalate rather than assume.

  • Policy violation

    Block execution and route to review.

  • Large blast radius

    Require explicit approval.

Start with the system. Tell us what needs to work.

An infrastructure decision, a migration, or a workflow you want AI to help with. Bring the problem, the data you have and the outcome you need. You do not need to choose a model or framework first.

Discuss your use case