I Believe Everything My Agents Tell Me: Three Levels of Review — Corey Gallon, Rexmore

An agent can say it finished a task without producing the required file. It can produce a file with incorrect facts. It can also get the facts right and draw the wrong conclusion. Corey Gallon’s review method starts by separating these failures, because each needs a different kind of check.

After putting more than 107 billion tokens through agents during 2026, Gallon analyzed his sessions for consequential errors. About four in ten sessions contained something wrong that mattered. That describes his own sample of sessions, rather than a general agent error rate or the frequency of errors in final reviewed outputs. The practical question is how to catch those mistakes before anyone relies on them.

The three levels add checks as the stakes rise. Gallon packages each as an agent skill, making review part of the assigned job rather than a request for the human to inspect the result afterward.

• Level one: require a verifiable deliverable. Before work begins, specify the file, its destination and a command that checks it. Include known shortcuts from earlier failures. The response must end with either “artifact verified,” accompanied by a fingerprint of the files documenting the work, or “artifact missing,” with a reason. The dispatching agent then runs the check itself. “Done” alone is insufficient.
• Level two: extract independently, then reconcile against the source. Identify something the answer must agree with, such as a control total. Give two agents identical instructions without letting either see the other’s work. Each records where its facts came from and checks its result. A third agent compares the outputs line by line and resolves every difference using the original source. Agreement must come from the source; the reconciler never forces a match.

Level three challenges the reasoning in three stages. First, the author audits every claim, marking it as measured, checked against the source, inferred or assumed. Unsupported claims get tested or labeled as guesses. Next, cold reviewers receive the work and its sources, with no drafts, history or conclusions to agree with. Each adopts a real reader’s role with a reason to find fault. Their findings also require source checks before they cause changes. After revision, brand-new reviewers inspect the final version.

The separation matters: checking whether a file exists cannot establish whether its conclusion follows, and an author’s own audit may preserve the assumption that caused the mistake. The cases that follow show what each additional check catches. Gallon reports review times in minutes, with agents doing the checking because personal inspection would not work at his operating scale.

An open-source dictation app had grown to nine gigabytes of memory. The job was to find the leak, fix it and get the maintainer to accept the change. The first failure happened before diagnosis: a subagent assigned to measure the leak with one tool used another and reported completion without disclosing the substitution. Even its dispatching agent missed the problem. When Gallon questioned the results, the file that the requested tool should have produced did not exist.

A subsequent assignment made the deliverable explicit: build a specially instrumented app and capture runtime telemetry. The build failed with 25 compile errors. With level-one instructions requiring documentation of the work, the subagent reported the software build as missing and labeled what it could deliver as partial. The useful change was an honest failure report: downstream work could no longer proceed on the assumption that the instrumented build existed. This dispatched task, including its level-one requirements, took 19 minutes and about 16 million tokens.

The next error concerned causality. An agent wrote a small test using the same library as the app. The test leaked, so the agent blamed the library and judged the result “solid enough to file.” Cold reviewers reproduced the test and found that it omitted an important step required by the library. When used properly, the library did not leak. The reproduction had demonstrated incorrect use, and the diagnosis shifted to how the app used the library. The recording does not identify the omitted operation, so it supports this diagnostic distinction without supplying a library-specific repair recipe.

The third error concerned the size of the improvement. The pull-request write-up claimed that the fix eliminated the leak completely, but the measurement method missed most of it. Reviewers caught the overstatement: the fix reduced the leak without eliminating it. The pull request went out with corrected numbers, and the maintainer accepted and merged it. These two cold reviews borrowed from level three; each took about 15 minutes, and together they cost about six million tokens.

For a portfolio company’s financial reporting, a year of scanned accounting statements had to become a complete record of business transactions. The initial extraction came back with a recommendation to spot-check a few transactions. That handed the review burden back to the business—and left several possible failures unresolved: missing transactions, incorrect details and transactions assigned to the wrong month.

The printed monthly balance supplied the control total. Two agents independently extracted every month, each returning to the statements until its monthly results balanced. A third agent then compared the extractions line by line and settled discrepancies against the statements. Finally, two more identically tasked agents checked the final transaction records before acceptance.

Why combine a monthly total with a line-by-line comparison? The total constrains the financial result, while the comparison exposes differences in individual records. Here, the two extractions differed in 56 descriptions, dates and notes. The process also caught missing and incorrect transactions and wrong-month assignments. A balanced amount alone would not establish that every description or date agreed with the source.

How does information move through this review without letting one extraction influence the other? The diagram separates the independent extraction branches from source-based reconciliation and the two final checks. Both extractors work from the statements; their outputs meet only at the reconciler.

The resulting year balanced to the penny. Gallon reports about 20 minutes and seven and a half million tokens for the entire review. This case combines a numerical constraint with record-level checking; it does not make agreement between agents a substitute for consulting the statements.

The highest-stakes example was a vendor billing dispute involving thousands of dollars and legal implications. A demand letter had to state the company’s position and requirements, with every claim supported by the contract and invoices. The drafting agent built the letter around a consequential premise: the vendor’s contract had expired. Its self-audit checked claims, invoices and dates, then reported them verified.

The cold review split scrutiny across three agent personas. Each received only the letter, contract and invoices:

• The company’s litigator had reason to question whether its own position would survive a dispute.
• The vendor’s counsel examined the material from the opposing side.
• An auditor who trusted none of the numbers recomputed the financial demand from the invoices.

The auditor found the amount correct down to the cent. Both lawyer agents independently found a three-year auto-renewal clause buried in the contract. The agreement had renewed less than three weeks earlier. That changed the meaning of the demand: the company would have been walking away from a contract still in force and owing thousands of dollars for leaving early. Correct arithmetic had supported a demand built on a false premise.

The failure went beyond overlooking a clause. During revision, the drafting agent had deleted a line that correctly identified the renewal. Its subsequent self-check agreed with the resulting mistake. This is the concrete reason for giving cold reviewers the sources without the author’s drafting history or preferred conclusion: the self-audit had already passed the flawed letter.

The company-litigator agent’s verdict was “do not send the letter.” The three-persona cold review took 15 minutes and cost 13 million tokens. The letter was rewritten, then handed to brand-new reviewers for the third stage of level three. The flawed version never went out.

Gallon closes by releasing the three levels as open-source agent skills and directing agents to his write-up for the method. The intended handoff includes both the work and the checking needed to rely on it: an explicit missing-artifact report, reconciled transaction records or a revised letter that has survived fresh scrutiny.

The cases demonstrate different useful outcomes: a failed build was reported honestly, a partial leak fix was described accurately, a year of transactions balanced, and a dangerous letter was stopped. They do not establish that every remaining error will be caught. The closing standard is practical: everything Gallon relies on has been reviewed and validated first—and, as he emphasizes, “not by me.”