--- title: "Generator and Evaluator: Why an Agent Should Never Review Its Own Work" description: "The generator and evaluator pattern separates who produces from who reviews, so your AI agents catch errors before your clients do and deliver better work." url: https://addabra.com/en/blog/generator-and-evaluator-agents lang: en date: 2026-08-30 category: "Harness engineering" keyword: "generator and evaluator" publisher: Addabra alternate: https://addabra.com/blog/generador-evaluador --- # Generator and Evaluator: Why an Agent Should Never Review Its Own Work > The generator and evaluator pattern separates who produces from who reviews, so your AI agents catch errors before your clients do and deliver better work. Thursday, 7:30 p.m. A finance manager is wrapping up the monthly board report with help from an AI agent. Before sending it, she types: "Review the report and tell me if there are any errors." The agent replies in seconds: "The report is consistent and ready to send." She sends it. The next morning, a board member asks why the margin on one business line jumped so much. The answer: the agent had added a figure in US dollars to another in local currency. When it reviewed the report, it made exactly the same mistake it made while writing it. This is not a rare glitch. It is what happens when you do not separate **generator and evaluator**. ## Why an agent can't review itself When you ask a model to review what it just produced, you are asking it to find errors in reasoning that seemed correct a minute ago. It tends to confirm it. Not out of bad faith, but for the same reason a person misses their own typos: they read what they meant to say, not what they actually said. Companies solved this long ago: the person who prepares a payment is not the one who authorizes it, and the person who drafts a contract is not the one who approves it. Segregation of duties exists because it works. Agents need the same principle. ## The generator and evaluator pattern, explained In March 2026, Anthropic published a guide on harness design for long-running agents. It describes separate planner, generator and evaluator roles, with verification gates between them. The core idea is simple: - The **generator** produces the work: a report, a proposal, copy, an analysis. - The **evaluator** reviews it against explicit, verifiable criteria, without having taken part in producing it. - If the evaluator finds problems, it sends the work back **with specific feedback**, not a vague "make it better." - The cycle repeats until the work passes the gate, or until it hits a cap on attempts, after which it is escalated to a person. Thoughtworks calls the things a harness gives an agent so it can correct itself "sensors." The evaluator is one of the most powerful sensors: it turns "I think it's fine" into an actual check. ### What a good evaluator checks An evaluator should not offer opinions in the abstract. It needs a list of criteria, for example: 1. Are all figures in the same currency, and do they tie back to the source? 2. Is every claim backed by data in the company brain? 3. Are business rules respected, such as authorized discounts or legal deadlines? 4. Do the tone and format follow the brand guide? 5. Does it add anything beyond the obvious? That list is the foundation of evals, the tests that continuously measure an agent's quality. We cover them in [evals for AI agents in production](https://addabra.com/en/blog/evals-for-ai-agents-in-production). ## Two reinforcements: cross-auditor and baseline agent ### A second model, from another provider, as cross-auditor If the generator and the evaluator are the same model, they share blind spots. What seemed reasonable to one will seem reasonable to the other. That is why, for high-impact tasks, the evaluator should be a model from **a different provider**. If the generator uses an OpenAI model, the auditor can use Claude or Gemini, or the other way around. This is not about one model being better than another. It is about them reasoning differently. When two models trained in different ways agree that something is correct, confidence goes up. When they disagree, that disagreement is exactly the point a person should look at. ### A baseline agent that writes the obvious The second reinforcement is less well known and very useful. Before the generator gets to work, a **baseline agent** produces the obvious version: the answer anyone would give using public information and no knowledge of your business. That obvious version sets the floor. The evaluator compares the generator's work against it and rejects anything that does not beat it. That way nobody, neither the agent nor your team, wastes time repeating what is already known. In a sales proposal, for example, the baseline agent writes "we offer quality, experience and great service." The generator is forced to go further: concrete cases from the client's industry, the specific problem they raised in the meeting, the commercial terms that set you apart. We apply this in [sales proposals in minutes](https://addabra.com/en/blog/sales-proposals-with-ai). ## Illustrative example: the board report, with the full pattern Back to the report from the opening. **Before:** a single agent drafts and "reviews." The currency error slips through. The finance manager finds out in front of the board. **After:** 1. The **baseline agent** writes the obvious summary: "sales grew, costs held steady." 2. The **generator** produces the report using figures from the brain, explaining the drivers behind each variance. 3. The **evaluator**, running a model from a different provider, checks it against the list: it catches that one figure is in dollars and another in local currency, and that the paragraph on costs does not beat the obvious version. 4. The generator fixes both issues based on the feedback it received. 5. The report passes the gate and reaches the finance manager, who reviews it and signs off before it goes to the board. None of this required a more expensive or smarter model. It required separating roles, writing clear criteria and putting a gate before the sign-off. It is organization, not magic. The finance manager does not spend her evening hunting for errors. She spends it asking whether the conclusions are right, which is her real job. The final human sign-off is the last gate; we explain how to design it in [human approval by role](https://addabra.com/en/blog/human-approval-by-role). ## Common mistakes 1. **Asking "review your work" in the same conversation.** It is the same agent with the same context: that is not a review. 2. **Vague criteria.** "Make sure it's right" is not a criterion. "All figures in the same currency" is. 3. **Loops with no cap.** Without a maximum number of attempts, generator and evaluator can argue indefinitely and drive up costs. 4. **Dropping the human sign-off because "there's already an evaluator."** The evaluator reduces errors; accountability still sits with a person. ## What to do on Monday Identify one deliverable that AI produces today and that someone sends out without an independent review. Write five verifiable criteria for that deliverable. With that list, you already have the core of an evaluator, even if a person applies it at first. At Addabra, this pattern is built into our harness: we separate generation from evaluation, use models from different providers as cross-auditors on critical tasks and measure everything against a baseline version. So whatever reaches your desk for sign-off has already passed through several gates. ## Frequently asked questions ### What is the generator and evaluator pattern? It is a way of organizing agents in which one produces the work and a different one reviews it against explicit criteria. If the evaluator finds problems, it sends the work back with specific feedback and the generator fixes it before it reaches a person. ### Why use a model from a different provider as the evaluator? Because a model tends to share the same blind spots with itself: if a line of reasoning seemed right when it wrote it, it will probably seem right when it reviews it. A model trained by a different provider reasons differently and catches errors the first one misses. ### Does the evaluator replace human review? No. The evaluator filters out verifiable errors before they reach a person, so whoever signs off spends their time on judgment rather than hunting for typos. The final decision on what goes to the client still belongs to a person in the right role. ## About Addabra Addabra is an AI-software-first software factory that designs and builds custom agentic systems for companies: one agent per process, a shared business memory, orchestration, human approval by role, auditing and cost control. Free diagnosis: https://addabra.com/en/blog/generator-and-evaluator-agents#contacto