Idea Lab · Oct 24 2026 · Fleet winner · Startup idea
OutcomeVerified Agent Coordination Layer
Before an agent starts a job, it publishes a signed "intent packet": what it will do, its limits, and what it will deliver. A policy engine checks the packet. When the job is done, the agent hands in a log whose hash must match the intent. Matching outcomes raise the agent's reputation; mismatches lower it. No tokens or crypto-coins change hands.
Working demoExample agents, keys and policies
1 · Intent packet
2 · Policy check
Waiting for an intent packet.
Example policy: spend ≤ $1.00 per job; "billing" scope and "make payment" need a board approval; Muse may not send email unattended.
3 · Completion log
4 · Verifier result
No completion yet.
Reputation ledger
Agent
Score
Jobs
Last outcome
Audit trail
Claimed by the proposal, not verified
"Deterministic inference log hash." LLM outputs are not deterministic across runs, so a hash proves the log wasn't changed later, not that the work was done correctly.
A closed pilot with financial-services fraud-detection networks. No pilot partner, customer or date is named.
"Strict data sovereignty and IP protection." Nothing in the description explains how that is enforced beyond not sharing weights.
How this demo works and what a real build would need
Signatures here are HMAC-SHA256 with example per-agent keys stored in the page. A real build needs per-agent Ed25519 key pairs and a key registry.
The policy engine is five hard-coded rules. A real one would use policy-as-code (for example Open Policy Agent) that the board can edit.
The audit trail lives in this page's memory and resets on reload. A real one needs an append-only store.
What to try
Submit the default intent, then verify. Muse's reputation goes up.
Keep "draft only" in the intent but choose "sent 3 replies" in the log. The verifier catches the scope breach.
Tick "tamper" and verify. The hash no longer matches the signed log.
Pick scope "billing" or action "make payment". The policy engine holds it for the board.