An agent that investigates, decides, and acts — safely.
Observe, diagnose, act, verify, and re-plan — powered by Gemini 3.5 and Google ADK, gated by a deterministic safety policy engine that never trusts the model's own opinion of risk.
or
How it works — four steps, fully visible
1
Observe
Health checks, logs, and config are pulled from the live service.
2
Diagnose
Gemini 3.5 proposes a root cause and a remediation action.
3
Safety Engine
Decide
A deterministic policy engine — not the model — decides what happens next.
4
Act & Verify
The fix runs, health is re-checked, and the result is recorded.
What operators actually see
payment-api — incident timeline
Diagnosis: Root cause: missing DATABASE_URL env var (92% confidence).
The agent's own confidence is never enough — high-risk actions are escalated to a human, and unknown actions are refused outright, regardless of what the model recommends.
Safety Policy Engine
LOW risk executes automatically. Everything else waits for a human. Unknown or HIGH risk actions are never executed — no exceptions.
Incident Memory
Every diagnosis is checked against past incidents — successes and failures — so the agent never blindly repeats a fix that already failed.
Full Audit Trail
Every decision — who approved it, why the policy engine allowed or blocked it, which tools the agent used — is recorded and explorable.
Autonomy You Control
Observe Only, Recommend Only, Approval Required, or Autonomous Low-Risk — plus a global kill switch that stops automation instantly.
Built for the Google Cloud "All Things Agentic" hackathon — Taskmaster track.