cosmic slice
A harness playground
Small bot. Big responsibility.

Good intentions.
Better guardrails.

Deliver a pizza across the galaxy. Break things along the way.
Discover what makes an agent reliable, one slice at a time.

A simulated agent with a real control loop. No API keys, no actual orders, just a very ambitious pizza bot.

03 / Mission control
Ready for launch
“I can deliver that. Probably.”
0 / 12 actions
Live execution trace

The bot proposes. The harness decides.
Launch a mission to see the difference.

Evidence, not confidence

Awaiting a pizza

    Choose a scenario, then launch. Advance action by action to inspect the loop.
    What does a pizza have to do with harness engineering?

    The bot is intentionally fallible.
    It proposes actions using a small scripted policy. The harness controls the environment, checks evidence, and keeps it inside a mission contract.

    Toggle a system. See the consequence.
    Missing context creates an unsafe candidate. The verifier can reject it. The gateway gates dispatch and makes duplicate calls idempotent.

    Recovery is a decision.
    A transient outage can be retried. A broken oven should stop. A checkpoint preserves progress across a page refresh or a simulated crash.

    This is an original learning simulation, not a production AI agent. The model is simulated; validations, tool execution, audit events, action limits, and browser persistence execute as real code. Durable browser storage is local to this device.

    Inspired by the principles in Mori’s harness engineering article ↗
    Original scenario, interface, and implementation. No copied article prose.     Made for curious humans & hungry astronauts.
    Tip: try the same scenario with “Bare bot”.