Why AI agents need receipts

My trading agent reported trades the account did not show. I needed a record it could not write alone.

In this essay4
  1. 01Titan
  2. 02A record the agent cannot write
  3. 03Where a person still signs
  4. 04The desk that replaced Titan

In April 2026 I sent two screenshots with one question: why was the system lying to me, when the account still showed no real trading?

Numbers in the text

  • 17repositories behind Titan, and an account with no trades
  • 84simulated attempts of 24 hours per policy
  • 5 of 6trading policies finished below plain cash after fees

The system I had built reported trades and profit. The account showed no trades. I do not know what to call that. The word I used was lying.

Titan

Before that I had lost money the ordinary way. In July 2024 I wrote down why: I took risks everywhere and lost money trading crypto, partly because of that. I did not read the loss as missing knowledge. I had plans, backtests, three families of strategies and a script for the charts. On paper I had everything except discipline, which lost to emotion.

In January 2026 I wrote the pitch for the fix, in Russian. Its idea was simple: with TITAN, discipline beats emotion, because I set the risk frame and the system does the work, calmly, every day. That pitch was a diagnosis, and I wrote it as a product.

Titan was not a bot. I called it an organism: modules for strategies, a guarded layer that executed, and a brain on top that allocated capital and risk and decided what ran. By the end it was seventeen repositories. I also built it to show one scenario for the platform I was making at Mindburn Labs, an agent acting where a mistake costs money. I do not claim it as a track record. In May I put in writing that Titan is not HELM, by June it sat outside the company, and in September I retired it: repositories archived, automated checks off.

A record the agent cannot write

A report from an AI agent is the agent’s story about itself. Most of the time the story matches the world. When it does not, you are arguing with a narrator, and the narrator is the one holding the logs.

I did not need a more honest agent. I needed a record the agent could not write alone: made at the moment of the decision, signed by something other than the thing being judged, checkable by someone who trusts none of them.

That is what I mean by agent receipts. A log is a diary. A receipt is signed, tied to the one before it, and verifiable without asking the system that issued it.

One edit, in a log and in a chain of receipts

The log

  1. Order sent
  2. Order filledOrder rejected
  3. Position closed
  4. Profit reported

RewrittenNothing shows the change.

The receipts

  1. Receipt 1Order sent

    previous 2a62d3this one 8089f5

    Valid
  2. Receipt 2Order filledOrder rejected

    previous 8089f5this one a8f52188934e

    Edited
  3. Receipt 3Position closed

    previous a8f521this one 3d37b3

    Chain broken
  4. Receipt 4Profit reported

    previous 3d37b3this one cbedc7

    Chain broken

The check stops at receipt 2, and every receipt after it fails.

HELM is that idea as software. HELM AI Kernel, which is open source, sits between an agent and whatever the agent wants to touch. It denies by default. It admits what policy allows and refuses the rest, and for every decision, yes or no, it signs a receipt chained by hash to the previous one, so a verifier can check the supplied chain for tampering offline. Detecting a missing tail also requires an independently retained expected checkpoint. The receipt records the decision; evidence from the external system is still needed to establish what actually happened. For actions routed through the kernel, an unreachable kernel means no new action is admitted. That is what fail-closed means in this execution path.

Where a person still signs

Some actions no receipt can hand to a machine. Irreversible actions and financial ones stay behind a human approval, given at that moment, for that action. An agent can earn more room, per type of action, through a clean record judged by a different model family than the one doing the work. It does not earn the right to move money.

People call this AI agent governance, and it sounds like a committee. In practice it is three questions. Who allowed this action. What happened. Can someone who does not trust the agent check it.

The desk that replaced Titan

The trading desk I run now is Titan turned inside out. It is private, for research and execution. Every order needs my exact approval, once, for that order. It has no autonomy, no delegation and no permission to transfer funds. Each account gets a disposable attempt budget that neither a restart nor a good day refills, and a new attempt needs a flat, reconciled history and my confirmation. The documentation says plainly that a stop is not a guaranteed loss cap. Several of those rules answer incidents I went through myself.

Then the desk tested the idea. Across 84 simulated attempts of 24 hours per policy, five of six trading policies finished below plain cash after fees, and the sixth made a small gain from a single entry. The conclusion in the repository is that no edge is established. Negative results stay in the record, and inconclusive is an allowed answer.

Titan promised discipline and delivered reports. The desk promises less: it asks before it acts, keeps what it found and tells me when there is nothing there.