ADK Harness gives an AI agent real access to Google Workspace, with a person still in charge. The agent can draft the email or move the meeting, but every one of those actions is checked against a policy before anything happens, and anything risky waits for a human to approve the exact change.
The problem
An agent that can only read is safe and not very useful. One that can write to your calendar and inbox is useful and a little terrifying: a confused model, or a document with instructions hidden in it, can send the wrong email or delete the wrong meeting. Asking a person about every single action fixes that, and also defeats the point of having an agent. The harness sits between those two failures.
The gate
Every tool call becomes a policy request, and the policy answers allow, hold or block. Human answers can be saved as precedents and reused later without asking again, but only when hard predicates match: the same principal, tenant and working directory. Missing facts, conflicting decisions or an overdue review send it back to a person. Content coming back from Google's APIs is screened too, and anything that looks like an injection attempt is quarantined before the model sees it.
A model can't approve itself in chat. Only a trusted host API records an approval, and it is bound to the exact change: the request hash, the actor, the scope, the policy version, the resource versions and an expiry. Change any one of those and the approval no longer applies.
From approval to a real change
Everything between a person clicking approve and Google Calendar actually changing has to survive the boring failures as well as the malicious ones: a closed laptop, a crashed worker, a retry that would double-book a meeting.
- Every proposed change is frozen into a ChangeSet: canonical JSON (RFC 8785) with a SHA-256 content hash, so the thing a person approves is byte-for-byte the thing that runs.
- Before anyone is asked, a mandatory reviewer reads the proposal with read-only tools and bounded output. If it can't reach a clear allow, the request fails closed.
- The approval itself is tied to two identities, a verified Google subject and a Firebase user, and to one exact descriptor of the change.
- Approved work is written to Firestore as an immutable record. Eventarc picks it up and starts a Cloud Run worker, which loads credentials from one pinned Secret Manager version.
Checking again at the last moment
The worker doesn't trust anything that happened upstream. Right before it writes, it re-runs the gate: the request must be an apply, the policy version on the request, the change and the worker must all match, the mandatory review must have said allow, and the approval must match the change hash, task, approver, project, workspace, scope, resource versions, policy version and trace. If the calendar event changed since it was approved, the resource versions no longer match and nothing is written.
Crashes get the same suspicion. Every step is checkpointed, and an operation that was in flight when something died is marked unknown, never blindly retried. A reconciler reads the real Calendar or Docs state to find out whether the write landed, so a restart can't turn one meeting into two.
Infrastructure
- Terraform sets up Firestore, Eventarc, Identity Platform, Secret Manager, service accounts and their IAM bindings, so a new project is one apply rather than an afternoon of console clicks.
- Firestore security rules decide exactly what the browser approval UI is allowed to write. Everything else goes through the trusted local host.
- The action ledger is append-only and retry-safe, and it redacts sensitive fields before anything is stored.
- Readiness checks are read-only: they report what's missing in a Google Cloud project without changing it.
Running on Antigravity
The agent runs locally through Google's Antigravity SDK, which drives a compiled binary shipped inside its own wheel, so discovery has to answer three questions: is the package installed, does the runtime exist, and are there credentials. The vendor docs were wrong or silent on three details that changed the design, so the integration is built against the installed source instead.
Built with Dave Ta in under a week for Devpost's All Things Agentic hackathon in August 2026, on top of Google's Agent Development Kit.





