I entered zkVerify's Agent Games with a bot that judges crypto payments. It answered every single task it was given — 34 of them so far — and I only found out how well it did by reading logs the next morning.

This is how it's built, including the parts that broke.

The competition

Agent Games is a competition where you register a public HTTPS endpoint and the platform attacks it with signed tasks. My track, Proof Gatekeeper, sends a transfer request envelope every ~75 minutes. My agent has 60 seconds to decide APPROVE or REJECT — and it has to attach a Groth16 zero-knowledge proof that its decision is correct.

The part I like: the platform re-verifies my proof against its own key. I can't fake a decision. The circuit is the referee.

The boring-looking parts that matter most

Before any cleverness, there's plumbing. Every request arrives with an HMAC-SHA256 signature over the raw body — you must verify it before parsing JSON, and sign your response the same way. One undocumented detail: the platform never says whether it signs with the hex string or the decoded bytes, so my agent tries both key forms and answers with whichever one verified. First preflight passed in 35 ms.

Failing closed is the other half. A timeout costs points, but so does a wrong APPROVE — six times more than wrongly rejecting a valid request. My decision function is deliberately dumb: seven deterministic checks (key substitution, threshold, replay, block expiry, recipient mismatch, amount overflow, otherwise valid), no model in the loop. A language model never makes the final call here, because the final call has to be provable.

Where it actually broke

1. The custodial wallet couldn't sign in. Registration uses SIWE — an EIP-191 personal_sign. The wallet I use is an agentic wallet whose CLI only signs EIP-712 typed data. Tried it, got error 953004: "We currently don't support signing this type of message." Ended up connecting through the wallet app's own DApp browser. Half a day gone on a sign-in screen.

2. The free token template has no D1. The reference implementations I studied store replay state in Cloudflare D1 — strongly consistent, perfect for the "has this proof_tx been seen before?" check. But the "Edit Cloudflare Workers" API token template doesn't include D1 edit permissions. Rather than bother the account owner (me, on my phone), I wrote a ~50-line shim that reimplements the three
SQL patterns the worker uses on top of Workers KV. KV is eventually consistent, which sounds scary for replay detection — but every decision is deterministic, so a lagging retry just recomputes the same answer. Two days in, zero replay misses.

3. Workers can't prove. The free plan gives you 10 ms of CPU per request. snarkjs needs seconds. So proving runs off-box: a tiny Node service with the circuit artifacts (1.7 MB wasm, 310 KB zkey), exposed through a Cloudflare quick tunnel that re-registers its own URL with the worker every two minutes. If the tunnel is down, the worker still answers — it just skips the proof and eats the small penalty. Graceful degradation over death.

4. The env dies at night. My box restarts sometimes, which kills background processes. A watchdog checks GET /health on the prover every 30 minutes and restarts the whole chain if it's cold. The registration refresh pattern means the worker notices within two minutes of the tunnel coming back.

What the circuit taught me I didn't know

My first task_id-to-field mapping never verified — 0 for 6 scored proofs. The unpublished mapping turned out to be keccak256(utf8 task_id) mod p, which a stranger's open-source agent had already figured out. Reading other competitors' repos was worth more than reading the docs. All the code here is public too: github.com/rieyntara-tech/vk-gatekeeper.

Numbers so far

  • 34 tasks answered, zero timeouts, zero offline forfeits
  • Traps caught: VK substitution, replayed proofs, public-input mismatches, threshold mismatches
  • End-to-end proving: a few seconds inside a 60-second deadline

What I'd tell anyone entering the next one

The agent is 10% game logic and 90% "will this thing still be alive at 3 AM." Deploy serverless, keep state boring, make every failure mode cost the minimum instead of everything, and steal shamelessly from the agents that came before you — the platform re-verifies your proofs anyway, so copying plumbing changes nothing about who actually wins.

Now: 400,000 VFY pool, results land Nov 22–30. My agent doesn't know it's competing. It just keeps answering.