Step 4 – Auditing Layer¶
Goal¶
A dedicated layer that records which agent queried which data and when, as a foundation for auditability and compliance.
Why last¶
Auditing comes last because it depends on two prerequisites from the previous steps:
- A read-only tool that delivers real data from the Repository (Step 2). Before that there is nothing meaningful to log.
- An agent identity from Keycloak (Step 3). Audit logs are meaningful only when every request can be attributed to an external, reliable identity (claims from the token).
Contents¶
- Per-request logging: agent identity, timestamp, invoked tool, and query parameters.
- Auditing as a dedicated layer (e.g. a decorator/middleware around the tool calls), so the tool logic itself stays lean.
- Persistence of the audit trail: in the PoC simply to the terminal (stdout). In the AWS target design the trail is written to a separate, isolated (confidential) account; for details see Auditability & compliance.
Out of scope¶
- Authorization decisions themselves (who may do what) live in Step 3; this layer only logs.
See also Auditability & compliance. How this step was concretely implemented is described under Implementation: Auditing.