Personal project · Agent tools

Fleet Ops Lab

I built a small warehouse simulation to explore a question about agents. When a delivery looks stuck, what evidence would you want before acting on its advice?

Problem

Why this mattered

A missing acknowledgment can make a completed delivery look unfinished. Sending the robot again would be the wrong fix. I wanted a place to create that ambiguity on purpose and inspect what each service saw.

Build

What shipped

  • An interactive order-to-receipt demo. Place an order, drop the inventory reply, and watch receiving retry. The inventory ledger lets you check that the trays were counted once.
  • A blocked-aisle scenario with investigator tools to read fleet state, cite observations, and preview a reroute. The browser demo uses a deterministic investigator. The local version can give the same tools to a language model.

Stack

Tools and systems

GoWebAssemblyReactTypeScriptGraphQL

Decisions

Technical choices

  • The investigator can propose a route change. The operator reviews the preview and decides whether to apply it.
  • The simulator tracks the actual delivery after a route change. A successful command alone cannot close the incident.

Outcome

Result

You can replay a failure and trace a proposed fix back to the observation that justified it. After an approved reroute, you can follow the robot to the station and check the delivery receipt yourself.