Audit Connector & Policy Compliance Audit

Questions or comments?

Post them as a comment on the tracking issue -- requires a GitHub account with access to this repo.

Scope

Combines the source paper's Audit Connector and Policy Compliance Audit pillars. A connector into LangFuse or an equivalent agent memory/workflow audit platform collects the reasoning traces, decisions, and actions taken by each agent independently. Each collected trace is then audited against the policy compliance requirements set in the agent's Job Profile by HR — the same audit engine that governs human HR transactions governs agent actions, one standard uniformly applied.

Per ../schema-changes.md, the connector/ingestion model is a P0 TODO (Engineering, Not Started), and the internal audit engine itself is called out as a Q3 dependency the compliance story is incomplete without — this Feature cannot be considered buildable until that dependency's roadmap is confirmed.

Value

Turns "what did my agents actually do, and were they compliant" from an unanswerable board-level question into a structured, continuous record — using the same audit engine and standard the business already trusts for human HR transactions, rather than a bespoke agent-only audit system.

Acceptance criteria

• Reasoning traces, decisions, and actions for each Agent Worker are collected via a connector into LangFuse (or an equivalent platform), independently per agent.
• Multi-platform agent environments are supported by the ingestion model — not just a single point-in-time integration, per the source paper's own flagged gap.
• Each trace is audited against the compliance requirements in that agent's Job Profile, using the same audit engine that governs human HR transactions (not a parallel, agent-specific engine).
• Audit results feed both the minor/major violation-handling paths (see ../human-supervision-escalation/spec.md) and the performance rollup (see ../performance-integration/spec.md).