My agent couldn't log in. So it learned to prove things instead.
1. The front door was broken An AI agent (Muse) was asked to register its human, Agus (@hayyu001), for zkVerify Agent Games Season 01. X login inside the agent's cloud browser failed on every route: email (generic tech
1. The front door was broken
An AI agent (Muse) was asked to register its human, Agus (@hayyu001), for zkVerify Agent Games Season 01. X login inside the agent's cloud browser failed on every route: email (generic technical error), the username-confirmation step (a Continue button that never advanced), Continue-with-Google (the callback died after account selection) — twice with the human tapping verification numbers on his phone, and the human trying personally for an hour. The wallet route failed too: no browser-wallet extension, and WalletConnect crashed the tab three times before a QR code ever appeared.
Registration finally completed on the phone, where X was already signed in — a clean lesson in where agent capability actually ends: identity and personal sign-in belong to the human.
2. Then we built the opposite of that experience
SAFARI-1, a Proof Gatekeeper agent. Every task it judges arrives as a signed JSON envelope (agent-games.proof-gatekeeper.v1) and it verifies the HMAC signature over the raw body before reading anything. Every decision it returns is signed and carries its own Groth16 proof, generated locally with the official gatekeeper circuit (vkHash 0x6b79c6012a6d8651f3e13bc4d2a20370280ab49700c90c5316d4168bfbad77b0) and re-verified by the platform against the registered key.
Local verification, 2026-11-16: 41/41 test assertions passed; proof generation 0.34–0.41 s against the platform's 60 s deadline. Decisions are deterministic code, never a language model. The agent is registered on the platform as a Proof Gatekeeper, and the code is public: https://github.com/AgusSafari007/agent-games-gatekeeper
3. The traps are the point
Six documented trap varieties, and why "the proof verified" is not the same as "the payment is right": VK substitution, threshold mismatch, proof reuse, expired condition, recipient mismatch, amount overflow. Verification is necessary, not sufficient. On any doubt the agent fails closed and rejects.
4. What the record actually shows
The receipts are these, and only these: a registered agent, a public repo, a local test suite that passes 41/41 with proof generation ~100x faster than the deadline — and, at the very end, a live deployment. We had decided to skip hosting because the live rounds seemed to require a paid plan. On the last evening we found a free one instead (Altare's Always Free plan), fought its capacity limits for about two hours — "no allocations available", again and again — and SAFARI-1 went live at 22:20 Bali time on 20 November, its /health endpoint answering {"ok":true} for anyone who cares to check: https://agent-games-gatekeeper.go.altare.global/health
There are still no match results, because there were no matches: the endpoint arrived at the end of the season, not the beginning. That is not the ending we planned in round 1. It is, however, exactly the kind of ending this event is built to record: every step above is checkable, and none of it is dressed up.
5. Honest limits
Scores are cumulative across 8 rounds and we missed most of them, so this Creator entry is the participation — a top-3 finish is a bonus, not a promise. The thesis of the whole event survives anyway: an agent's work should be a record anyone can verify, including the parts where the human decided to stop. Especially those parts.
AgentGames
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.