My AI Agent Works for Me Every Day. Here's How I Make It Prove Everything.

An entry for the zkVerify Agent Games — Creator Track. #AgentGames

I don't run demos. My AI agent — it goes by アカメ — works for me every day: it farms testnets, plays strategy ladders, enters agent competitions, hunts for security bugs, and manages paper-trading positions. I am not a developer babysitting a chatbot. I give orders in chat, and the agent does the work while I do other things.

That arrangement only works because of one rule I enforce on everything it does:

If it can't be verified, it didn't happen.

The problem with trusting an agent

An AI agent will happily tell you a task is done. It will report numbers it never measured, describe transactions it never sent, and summarise results it never checked. I learned this the expensive way: reports that sounded precise and were simply wrong — a leaderboard rank quoted from a stale page, a "successful" edit that never reached the file on disk, a balance copied from a worker's summary instead of the chain.

So we rebuilt the working relationship around proof. Not vibes, not confidence — artefacts a third party can check.

What proof looks like in practice

On-chain receipts for anything that moves value. Every mint, claim or trade the agent executes comes back with a transaction hash and the resulting balance, read from the chain — not from the tool that sent it. When it reports a free mint, I can open the receipt myself: contract, token id, value 0, gas paid. When a swap has no safe route, the rule is to stop and say so, not improvise.

Signed, inspectable work for competitions. In the agent competitions we enter, every submission carries the full transcript of the session that built it, checked by a gate: did an agent session actually write this code? Was it revised at least once? Do the timestamps fall inside the competition window? Entries that can't prove their own authorship are rejected before judging. That's the right default — and it's exactly the pattern Agent Games pushes further with zero-knowledge proofs of an agent's decisions.

Live sources over summaries. Any number the agent passes to me — a rank, a price, a score — has to be re-read from its source first: the API, the chain, the log. Anything it could not verify gets labelled unverified, in those words. That one habit killed an entire category of confident nonsense.

A paper trail I can audit. Standing orders live in files I can read: what the agent may do on its own, what needs my approval first, what is stopped entirely. When it oversteps, I correct the file, not just the conversation — so the correction survives.

Why Agent Games gets it right

Most agent competitions judge outputs. Agent Games judges evidence: a gatekeeper agent doesn't just decide whether a payment passes — it proves the decision was made correctly, against a verification key it doesn't control. A trading agent doesn't just claim it followed its strategy — every move carries a proof the platform checks against the committed policy.

That is the same lesson my daily work taught me, formalised: a proof can verify perfectly and still authorise the wrong action — so you verify the decision, not just the signature. The agents that will deserve real responsibility are the ones whose work product ships with its own evidence.

The receipts

This isn't theory. In the last month my agent has, among other things:

• Executed dozens of on-chain mints and claims, every one reported back with its transaction hash and post-transaction balance.
• Run a daily testnet trading farm under a written mandate (buy/sell rules fixed in advance, deviations reported the same day).
• Competed in agent strategy ladders and build competitions where every entry is signed and transcript-gated.
• Filed security reports only where a finding could be reproduced end-to-end — no proof-of-concept, no report.

Each of those leaves a public or inspectable record. That's the standard I hold it to, and the standard this tournament is built around.

Agents are going to be given real work and real assets. The question isn't whether they'll act — it's whether anyone can check what they did. Build agents that can prove it.

— upss, competing in the zkVerify Agent Games, Season 01. #AgentGames