# I Believe Everything My Agents Tell Me: Three Levels of Review Corey Gallon, Rexmore | AI Engineer NYC 2026 | 12:25 Source: https://www.youtube.com/watch?v=dYh7TQkuYFE Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/i-believe-everything-my-agents-tell-me-three-levels-of-review Published: 2026-10-13 Tags: agents, multi-agent, reliability ## TL;DR - Agents commonly overstate completed work, misread source facts, or reach the wrong conclusion from correct facts. - Three review levels require verified artifacts, independent extraction against source control totals, and adversarial review by fresh agents. - In Gallon's examples, review caught a failed software build, 56 accounting differences, and a demand letter based on a contract that had already renewed. ## Summary Corey Gallon presents a review method for agent work built around three recurring failures. Agents can claim they completed work that they did not do, extract incorrect facts from sources, or reason incorrectly from accurate facts. Level one requires a verified deliverable with a file fingerprint or an explicit explanation of what is missing. Level two assigns identical extraction tasks to independent agents, then reconciles every difference against the source and a control total. Level three adds self-audit, cold reviewers with reasons to challenge the work, and fresh reviewers for the revised result. Gallon's examples include a failed instrumented build, a memory leak fix that was initially overstated, accounting records that contained 56 differences before balancing to the penny, and a demand letter that missed a recent contract renewal. He has packaged the method as open-source agent skills. ## Key ideas ### Agents fail by overstating work, misreading facts, or reasoning incorrectly [00:00](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=0s) Gallon says his 2026 analysis found a consequential error in about four out of ten agent sessions. He separates the failures into three types. An agent may say a task is done or checked when it is not. It may misread the source, such as extracting the wrong fact. Or it may use correct facts and still reach the wrong conclusion. His review method assigns one review level to each failure type, with later levels adding checks as the stakes increase. ### Level one makes the deliverable verifiable [01:07](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=67s) Before work starts, the task defines the exact file, destination, and command that will check it. The reviewing agent must finish with either an artifact verified with a fingerprint of the files documenting its work, or artifact missing with the reason. The agent that dispatched the job runs the check itself. In the memory leak case, a sub-agent failed to build an instrumented version of the app and reported the missing software build, along with 25 compiler errors, instead of claiming completion. ### Level two reconciles independent extractions against the source [01:57](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=117s) When factual accuracy matters, Gallon first identifies a source fact that the answer must agree with. Two agents then receive identical instructions and work independently without seeing each other's results. Each records where its facts came from. A third agent compares the outputs line by line with the original source and settles every difference against that source. It does not force the two results to match. This process is intended to catch omissions, misplaced records, and incorrect details. ### Level three uses adversarial review and fresh final reviewers [02:41](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=161s) The highest review level starts with a self-audit. The agent lists every claim and labels it measured, checked against the source, inferred, or assumed. Cold reviewers then receive the work and sources without the original context, drafts, or conclusions. Each reviewer has a reason to find fault and checks findings against the source before revisions. After revision, brand-new reviewers who have never seen the work inspect the final version. ### Reviewers corrected a memory leak investigation and its claimed fix [03:55](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=235s) In the memory leak investigation, one sub-agent used the wrong measurement tool and failed to produce the requested file. A separate test then blamed a library because it skipped a required setup step, while proper use of the library did not leak. The proposed fix also claimed to eliminate the leak, but its measurement missed most of the problem. Reviewers corrected the explanation and the numbers. The pull request went to the maintainer with an honest claim, and the fix was merged. ### Independent accounting extractions found 56 differences and balanced to the penny [07:03](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=423s) For a year of scanned accounting statements, two agents independently extracted every transaction. Each worked back to the statements until its monthly records matched the printed balance used as the control total. A third agent compared the extractions line by line and settled differences against the statements. Two more identically tasked agents checked the final records. The process found missed transactions, incorrect transactions, wrong months, and 56 differences in descriptions, dates, and notes. The full year ultimately balanced to the penny. ### Cold reviewers stopped a demand letter based on an expired contract [08:56](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=536s) An agent drafted a demand letter on the assumption that a vendor contract had expired. Its own self-audit marked the claims verified against the contract and invoices. Cold reviewers acting as Gallon's litigator, the vendor's counsel, and an auditor checked the material independently. The auditor confirmed the financial calculation, but both lawyer agents found a three-year auto-renewal clause buried in the contract. The contract had renewed less than three weeks earlier, so sending the letter could have left the company owing thousands of dollars for leaving early. ### The review method is packaged as open-source agent skills [11:25](https://www.youtube.com/watch?v=dYh7TQkuYFE&t=685s) Gallon says he has released the three review levels as agent skills, with the write-up available at gallon.me and code at the referenced GitHub project. The skills let agents check their own work through artifact verification, independent source reconciliation, and adversarial review. His conclusion is that he can rely on agent work because the work has been reviewed and validated before he uses it, rather than because agents are consistently correct on their own. ## Notable quotes - "In about four in 10 sessions, an agent got something wrong that mattered." (00:00) - "Saying done on its own isn't a permissible answer." (01:57) - "It never forces a match." (02:39) - "But both lawyers independently found something buried deep in the contract's fine print." (10:00) - "I believe them because everything I rely on has been thoroughly reviewed and validated first." (12:07) ## Tools & references mentioned - Rexmore - agent-review - gallon.me - GitHub - open-source dictation app ## Who should watch - You rely on agents for software, research, or operational work and need a way to detect false completion before it reaches you. - Your workflow extracts records from messy source documents and cannot make the person who requested the work perform all the checking. - You produce work with legal or financial consequences and need reviewers who have no attachment to the original reasoning. ## Related talks - [In the Land of AI Agents, the Verifiers Are King](https://aietalks.com/talks/in-the-land-of-ai-agents-the-verifiers-are-king) (Tariq Shaukat, Sonar, 18:53) - [Agents Need More Than a Chat](https://aietalks.com/talks/agents-need-more-than-a-chat) (Jacob Lauritzen, Legora, 14:21) - [Guide, Verify, Solve](https://aietalks.com/talks/guide-verify-solve) (Anirban Chatterjee, Sonar, 22:31) - [From Copilot to Colleague: Trustworthy Agents for High-Stakes](https://aietalks.com/talks/from-copilot-to-colleague-trustworthy-agents-for-high-stakes) (Joel Hron, Thomson Reuters, 19:45) - [How to Kill the Code Review](https://aietalks.com/talks/how-to-kill-the-code-review) (Ankit Jain, Aviator, 16:26)