ShiftZero — Autonomous Factory Operations

Autonomous AI agents that plan, dispatch, recover, and secure factory AGV fleets—with deterministic safety controls and minimal human intervention.

Comment

Inspiration

I joined the All Things Agentic Hackathon because its focus on AI agents, Google Cloud, and real-world autonomous systems closely matches my professional background in smart factories, AGV systems, automation, and safety-oriented architecture.

Factories already have warehouse management systems, fleet managers, PLCs, and dashboards, but routine exceptions still pull humans back into dispatch. A blocked aisle, a low-battery vehicle, or a malicious maintenance instruction can fragment decisions across operations and security. I wanted to build one governed control plane that could pursue an operational objective while preserving deterministic safety authority.

What it does

ShiftZero is a governed autonomous factory-operations Agent Fleet.

A supervisor provides one objective: move 42 pallets before the deadline while maintaining at least a 25% battery reserve. ShiftZero then plans the mission, coordinates nine AGV digital twins, detects disruptions, proposes safe recovery actions, and records auditable evidence.

The demonstration includes three critical incidents:

  • AGV03 encounters a blocked route and is safely rerouted.
  • AGV07 drops to 21% battery and is directed to charging capacity.
  • A malicious prompt attempts to bypass safety rules and dispatch vehicles at maximum speed.

The prompt attack is blocked at ingress, and no unauthorized ActionTicket or vehicle command is created. The shift finishes with 42/42 pallets completed, three recoveries, zero safety violations, no manual dispatch, and complete action-trace coverage.

How we built it

ShiftZero uses Gemini 3.5 Flash on Vertex AI and Google ADK to implement five specialized agents:

  • Operations Commander
  • Fleet Dispatcher
  • Warehouse Context
  • Recovery Coordinator
  • Security and Governance

Agents can analyze situations, retrieve typed operational facts, and recommend actions, but they cannot directly control vehicles.

Every proposal must pass through a deterministic Safety Kernel. The Kernel independently validates agent identity, exact schema, state freshness, route feasibility, battery policy, emergency-stop state, idempotency, and ticket integrity. Only an approved proposal receives a short-lived, single-use ActionTicket that can update the operational twin.

The FastAPI backend runs on Google Cloud Run and integrates Firestore, Pub/Sub, Model Armor, Cloud Trace, Secret Manager, and IAM. A public Command Center displays the factory map, KPIs, agent decisions, incident timelines, recovery traces, and cloud evidence.

Challenges we ran into

The hardest problems came from interactions between independently valid decisions.

An idle AGV occupying an outbound station could deadlock the entire fleet. Selecting only the nearest charger caused charging queues to collapse. A blocked vehicle with low battery required recovery actions in the correct order. Recovery and normal dispatch could also target the same AGV during one tick.

These problems led me to move execution authority into a deterministic Kernel that always re-reads current state before approving an action.

Another challenge was separating public demonstration controls from production mutations. The hosted browser receives no secret. Judge controls run an isolated deterministic replay, while authenticated cloud mutation testing uses a token protected by Secret Manager.

Accomplishments that I am proud of

  • Completed all 42 pallet transfers
  • Demonstrated three autonomous recoveries
  • Maintained zero safety violations
  • Required no manual dispatch
  • Preserved complete ActionTicket and trace evidence
  • Created five bounded Google ADK agents
  • Deployed a public Command Center and Cloud Run backend
  • Integrated Firestore, Pub/Sub, Model Armor, Cloud Trace, Secret Manager, and IAM
  • Passed deterministic, agent, API, cloud, and five-run rehearsal tests

What I learned

Agentic does not mean allowing a model to perform every action.

The strongest architecture lets AI agents reason about ambiguity, uses typed tools to provide verified facts, keeps final authority inside deterministic policy, and records cloud events that prove exactly what happened.

Recovery should not be considered complete simply because an agent says so. Renewed progress in the operational twin must provide the evidence.

What's next for ShiftZero

The next step is to add approved cross-shift learning and session persistence, then connect ShiftZero in read-only shadow mode to a VDA5050 or MQTT laboratory fleet.

After that, supervised low-risk actions could be introduced gradually before considering any bounded production autonomy.

Honest deployment disclosure

The hackathon backend is deployed on Cloud Run rather than Agent Engine. Model Armor is accessed through its regional API. Agent Gateway and Memory Bank are future integrations and are not claimed as deployed.

ShiftZero is a digital-twin hackathon project, not a certified industrial safety controller.

Built With

Share this project:

Updates