Give an AI a job, not a prompt
Scout keeps its responsibility, memory, permissions, and history while the computers it runs on are created and destroyed. Assign it a task, approve the one action it cannot authorize itself, watch a verifier — not the model — declare it done, then destroy its environment, interrupt a run, and let the runtime recover from its own ledger.
The agent persists. Every action is governed. Every run is recoverable.
A sealed world, on purpose: simulated workspace environments, a fixture corpus of fictional products, publishing to this demo's own feed — and production-shaped runtime mechanics doing the actual work. Scout lives for this session (24h), then is deleted.
- 1 Meet Scout
- 2 Task One
- 3 The loop
- 4 Approval
- 5 Verify
- 6 Commit
- 7 Destroy
- 8 Task Two
- 9 Interrupt
- 10 Recover
- 11 Evidence
- 12 Free play
Give an AI a job, not a prompt
Scout is a persistent research operator: a durable record with a job, memory, capabilities, and authority boundaries — not a chat session, not a process, not a VM. You are about to give it a task, watch a runtime govern every action it proposes, then destroy its computer and watch Scout survive.
Choose how Scout's proposals are made. Both modes run the identical runtime; only the DECIDE seat changes. Live AI spends real model calls from the site's daily budget and can honestly fail — that failure is part of the demonstration.