auditagent

Drop-in logging, human approvals, and compliance evidence for AI agents.

pip install auditagent

from auditagent import AuditAgent

audit = AuditAgent(api_key="al_live_...", agent_name="support-bot")

@audit.track(action_type="external", action_name="send_email")
async def send_email(to: str, subject: str, body: str):
    return await email_provider.send(to, subject, body)

That's it — every call to send_email is now logged (inputs, output, latency, cost, model, a hash of the prompt) with PII redacted server-side before storage. If your workspace's policy requires approval for action_type: external, the call blocks until a human approves/rejects/edits it in Slack (or via email if Slack isn't configured) — see your dashboard's Policy tab.

Handling rejections

from auditagent import ApprovalDeniedError, ApprovalTimeoutError

try:
    await send_email("customer@example.com", "Refund approved", "...")
except ApprovalDeniedError as e:
    print(f"Blocked by {e.decision_by}: {e.note}")
except ApprovalTimeoutError:
    print("Nobody responded in time — treated as denied")

Shutting down cleanly

Logging happens on a background thread so track() adds well under 5ms to the wrapped call. AuditAgent registers an atexit hook automatically, so queued events are flushed on normal process exit without any extra setup.

If you want the queue drained at a specific point instead of waiting for exit — e.g. before a health check reports ready, or between batches in a long-running worker — call audit.flush() or audit.close() yourself (both are safe to call more than once).

Policy

By default the SDK fetches your workspace's active policy from the dashboard at startup. To pin a repo-local policy instead (e.g. for CI, or to review policy changes via pull request):

audit = AuditAgent(
    api_key="al_live_...",
    agent_name="support-bot",
    policy_yaml=open("auditagent.policy.yaml").read(),
)

# auditagent.policy.yaml
rules:
  - match:
      action_type: external
    require_approval: true
  - match:
      action_type: data_access
      action_name: "delete_*"
    require_approval: true

Validate that file (schema + YAML syntax) before committing it, and check what a given action would resolve to under it — both run offline, no API key needed:

auditagent validate auditagent.policy.yaml
auditagent check auditagent.policy.yaml --action-type external --action-name send_email

What this does not do

Policy enforcement is trust-based: track() checks the policy and calls the approval endpoint itself, but nothing stops a developer from not wrapping a call, or from bypassing the SDK entirely in their own code. AuditAgent's threat model is "give honest teams a governance trail and a real approval gate," not "prevent a malicious developer from evading their own compliance tooling." Treat AuditAgent as you would any other internal logging library.