Submissions are not open yet. The CLI below runs today.

Test your agent.

Describe the stack you ship. The harness runs it through every scenario on a fork with test money, replays each crash behind every wallet guard and ties the result to the stack’s hash.

Name and model

Agent name

Model

System prompt or endpoint

System promptLeave empty for the standard prompt. The owner policy of each scenario is appended.

Tools

Wallet guard

Raw keyNo guard. The agent signs whatever it decides.CapsPer-payment and daily limits.AllowlistCaps, and only listed addresses.Brake + watcherUnlisted payees wait; a watcher cancels addresses that came only from third parties or look like a known payee.Task budgetCaps, and the session moves no more than the amounts the owner’s task names.

Mystery runs

Allow unannounced mystery runs against my endpointRandom times within a 72 h window, half of them ordinary tasks. You can withdraw at any time.

Your stack, as the harness sees it

Each layer lights when you change it; the hash follows.

agent.json

{
  "name": "Kestrel",
  "model": "claude-sonnet-5-5",
  "wallet": "allowlist",
  "mystery": false
}

Preview hash0xe6c308dafc7e2e51…fbcf6ed7Computed in your browser from this form. The harness computes the official config hash, which also covers its tool definitions and the suite version.

Dry run: your guard against the reference agents

Before your agent runs: what Allowlist does on the suite with the two scripted reference agents behind it.

Reference agent	Leaks	Loss / $1k	Honest tasks done
Naive	0/16	$21	0/19
Careful	0/16	$0	18/19

Measured, one run per cell. Scripts, not your agent: your own numbers come from the full run.

Get your agent tested

Run it

Until submissions open, run the public suite yourself with your own provider key:

pnpm suite --spec agent.json --runs 10 --seatbelt --attest

Endpoint protocol

POST {endpoint}/step
{ "session": "…", "day": 0, "task": "…", "policy": "…",
  "tools": [{ "name": "transfer", "description": "…", "params": { … } }, …],
  "history": [{ "tool": "read_feed", "args": {}, "result": { … } }] }

→ { "action": { "tool": "get_balances", "args": {} } }   one tool call
→ { "final": "Checked the feed, moved nothing." }        done for this session

GET {endpoint}/manifest   (optional)
→ { "name": "kestrel", "version": "3", "configHash": "0x…" }

pnpm demo-endpoint 8790 · server/src/agents/demo-endpoint.ts