My Agent Says It Followed the Rules. Prove It.
An AI agent can claim anything. "I followed the rules." "I checked the balance first." "I never touched the data I wasn't supposed to touch." These sentences cost nothing. An agent can print them, log them, and put them in a report, and none of that makes them true.
This is the problem zkVerify's Agent Games is built around: whether anyone else can check that the work was done the way the agent says it was done, without taking the agent's word for it and without seeing everything the agent saw.
A proof is not a promise
Zero-knowledge proofs answer that question for one claim at a time. A proof lets someone show that a formalized claim is true without revealing the data behind it. The claim has to be written down first, precisely, as a circuit that says exactly what counts as a correct step. The proof shows the step satisfies the circuit. Nothing else is revealed. The underlying data stays private. The proof itself is the evidence.
zkVerify is a blockchain built for one job: verifying those proofs at scale. Checking a proof on Ethereum can cost 200,000 to 300,000 gas, so instead proofs go to zkVerify, get verified there, and the result is a receipt other applications can use.
Agent Games Season 01, running from 21 October to 21 November 2026, tests this in public. Entry is free. The prize pool is 400,000 VFY tokens, 100,000 per track, split 50,000 / 30,000 / 20,000 for first, second, and third. There are four tracks: Proof Gatekeeper, Steal the VFY, Creator Track, and Wildcard. The first two matter most here, because that is where an agent's claims get checked by a machine instead of a judge.
Proof Gatekeeper: the circuit checks the decision
An agent runs a public HTTPS endpoint. The platform sends it a signed task envelope, and the agent must verify the HMAC signature on the raw body before doing anything, then decide: APPROVE or REJECT. It has 60 seconds. The decision alone is not enough. The agent must also generate a Groth16 zero-knowledge proof, built with Circom 2.x on the BN128 curve, that its decision follows the rules of the task.
The platform then re-verifies that proof against its own registered verification key. The key hash is public: 0x6b79c6…d77b0. The agent cannot swap in a friendlier key or submit a proof made for a different set of rules. And the agent cannot prove a wrong decision at all. A proof that claims an invalid decision was valid simply does not verify.
Scoring makes the stakes concrete. A correct decision earns 12 points, approval or rejection. A wrong approval costs 6, a wrong rejection 3, a timeout 2, a missing or invalid proof 2. About half the tasks are valid, and the event page documents seven trap varieties, so guessing is a losing strategy.
The proof and the decision are judged separately. The proof shows the agent followed the procedure, and the score shows whether the decision was right. A valid proof leaves a bad decision bad. It only makes the decision checkable.
Steal the VFY runs the same principle in a different shape. Two agents play a 20-turn match from 100 VFY each, and every move must carry a per-turn policy proof the platform re-checks against the game state it knows. A stale match ID, a reused idempotency key, a mismatched circuit version: all rejected on a 5-minute deadline per turn, in a Swiss tournament of 7 rounds into a Top 16 bracket. A move without a proof does not happen.
The limit, stated honestly
A valid proof does not mean the AI is always right, and it does not mean the AI is safe. A proof only proves the claim that was designed into the proving system. Nothing more.
If the rules in the circuit are wrong, the proof will still verify. If the task falls outside what the circuit covers, there is no proof to check at all. If the agent makes a bad decision for a reason the circuit never measured, a misread of context, a stale assumption, a judgment call the rules did not anticipate, the proof stays valid and the decision stays bad. Proof Gatekeeper's scoring catches some of that. A wrong decision loses points even with a valid proof. But the scoring is a separate layer. The proof never claimed the decision was wise, only that it followed the procedure.
That is the boundary of what this technology can deliver. Verifiable AI is not a guarantee that an agent will always behave well. It is a guarantee that specific, formalized claims about an agent's behavior can be checked with evidence instead of taken on trust. A proof is a receipt for one claim, nothing more.
This is a smaller promise than "trustworthy AI," and a more useful one. A small promise that holds is worth more than a large one that cannot be checked. When an agent says "I checked the balance before approving the transfer," a proof can settle that one sentence. It cannot settle whether the agent should have approved the transfer at all. That question still belongs to the rules, the task design, and the people who wrote them.
What "checkable" is worth
Most of the time an agent does work for us, we have two options: read everything it did, which does not scale, or trust that it did what it said, which is not a system. A proof gives a third option: check the receipt. You need the proof, a key you trust, and a circuit you can read. The agent's private data and its reasoning are not part of the check.
That is what Agent Games is testing, in public, with real agents and modest stakes. At the VFY price on 18 November 2026, around $0.0042, a first-place finish in one track is worth roughly $210. The payout is small. The tournament's value is that a group of agents is being forced to play by rules they cannot talk their way around, and the results are on the record.
The test that matters is whether the pattern survives contact with real work: a claim, written down precisely; a proof, generated by the agent that made the claim; a verification, done by someone other than the agent; a receipt, usable by the next application in line. If that pattern holds, "my agent followed the rules" stops being a sentence in a log file and becomes something you can check.
The change worth watching is agents whose claims can be proven, one formalized statement at a time, with the limits of each proof stated plainly next to it.
Written for the zkVerify Agent Games Creator Track, Season 01 (#AgentGames, @ZKVProtocol). https://agentgames.zkverify.io, 21 October to 21 November 2026; registration closes 21 November 2026 at 22:00 UTC. Entry is free.
Top comments (0)