💡 Agent Activity Audit Log for Cross-Workflow Observability

[github-actions[bot]]

Summary

Implement a structured, queryable audit log that captures every significant decision and action taken by AI agents across all org workflows. Track what each agent read, modified, created, or skipped, along with the reasoning context. This bridges the gap between GitHub Actions execution logs (which capture that a step ran) and AI agent decision-making (which captures what and why).

Market Signal

GitHub's Actions Data Stream (public preview in 3-6 months) will deliver near-real-time execution telemetry to Amazon S3 and Azure, but covers workflow-level metrics, not agent-level decisions. The OWASP GenAI Exploit Round-up Q1 2026 recommends "full observability of agent actions and reasoning" as a key defense against supply chain attacks. Microsoft's security blog on AI agent RCE vulnerabilities emphasizes the need for audit trails of agent tool invocations. LLM monitoring tools (LangWatch, Braintrust) are converging on agent-level observability as table stakes for production AI systems.

User Signal

The org runs multiple autonomous agents (compliance-audit, feature-ideation, org-status, dependabot-automerge) that make decisions without direct human oversight. When these agents skip a finding, file an incorrect issue, or make unexpected changes, there is no centralized way to audit what happened and why. The daily-org-status workflow (PRs #200, #231, #234, #258) has been iteratively improved to capture outcomes, but it captures what happened, not why the agent decided to do it.

Technical Opportunity

Each agentic workflow can emit structured JSONL entries to a workflow artifact with a consistent schema:

{
  "timestamp": "2026-05-13T12:00:00Z",
  "workflow_name": "compliance-audit",
  "agent_name": "claude",
  "action_type": "create_issue",
  "target_resource": "petry-projects/broodly#42",
  "reasoning_summary": "Missing required workflow: agent-shield.yml",
  "tokens_used": 1234,
  "outcome": "success"
}

The structured logging standards in AGENTS.md already define snake_case field naming conventions and canonical field names (correlation_id, request_id). A lightweight log_agent_action() shell function in scripts/lib/agent-telemetry.sh provides the API. A weekly aggregation step in daily-org-status collects artifacts and produces an agent activity digest.

This aligns with the forthcoming Actions Data Stream by producing complementary agent-level telemetry that GitHub's stream does not cover.

Assessment

Dimension	Score	Rationale
Feasibility	med	Requires instrumentation across multiple workflows; schema design needs care to be useful
Impact	high	Enables security forensics, cost attribution, and compliance audit trails for all AI agent activity
Urgency	med	No immediate incident, but the org's agent fleet is growing and the audit gap widens with each new workflow

Adversarial Review

Strongest objection: GitHub Actions logs already capture workflow execution. Adding another logging layer creates maintenance burden and log sprawl for a small org.

Rebuttal: Actions logs capture that a step ran and its exit code, not what the AI decided within that step. When Claude skips a compliance finding, modifies a file unexpectedly, or generates a discussion, the Actions log shows "step completed" but not the reasoning. The Clinejection attack was discovered only after the malicious npm package was published — an agent-level audit trail would have caught the anomalous tool invocations. The implementation is a thin logging wrapper (JSONL output per workflow) with zero new infrastructure — no database, no service, just files. It also supports cost attribution (bridging to the Token Cost Observatory proposal) and compliance (providing the audit trail that SOC 2 reviewers expect for automated decision-making systems).

Suggested Next Step

Define the agent activity log schema (fields: timestamp, workflow_name, agent_name, action_type, target_resource, reasoning_summary, tokens_used, outcome). Add a log_agent_action() shell function to a new scripts/lib/agent-telemetry.sh. Instrument one workflow (compliance-audit) as a pilot. Review the output for a week before expanding to other workflows.