A governed institutional-agent fleet for continuous access review
Continuous access review needs more than an automated scan: it needs durable identity, minimized data, deterministic policy gates, and an agent fleet that remains unable to grant privilege.
Independent system · architecture, implementation, deployment, and evidence
The deployed evidence set demonstrates private, read-only access review across managed agents, durable state, event delivery, policy enforcement, redacted escalation, and payload-free audit records.
Access review is quarterly work performed on continuously changing permissions. Automating the scan isn't enough — an institutional agent must audit the real policy rather than fixture rows, survive asynchronous retries, prove why it acted, and remain unable to turn suspicious input into a privileged write.
Deterministic Python reads the live project policy, classifies broad roles, derives an HMAC-backed opaque finding ID and computes a bounded risk score. A compliance product cannot answer “why was this flagged?” with “the model thought so”, so Gemini explains already-minimized categories and decides nothing.
The auditor declares an output schema, so findings cross the A2A boundary as validated structured data: the finding ID must match a 24-hex pattern and the reason must be one of three deterministic codes. The scoring step downstream holds no model, so nothing retypes a finding on the way past.
A policy gate refuses to continue unless scoring left its own result behind, and an escalation gate refuses to report completion unless every routed department confirmed delivery. Both live in the orchestrator, not in the remote workers they check — a guard that travels over A2A is one the caller has to trust the callee to run.
The operational fleet is private by design. What is public is a sanitized evidence console, the architecture, the code and the recorded demonstration — deliberately not a route into a live IAM endpoint.
I’m always interested in the constraints that make an architecture worth explaining.