The author describes a three-week experiment testing AI agent trustworthiness, finding that agreement between agents using the same model is unreliable, and outlines a system with human approval for critical decisions.
I spent three weeks testing whether a team of AI agents could produce trustworthy work and not just more work. The biggest finding was that agreement between agents means very little when they are running the same model. I documented that failure, two experiments that produced no improvement, a hidden-permissions problem, and an agent that safely handled 21 order-desk calls inside a database enforced lane. With the AI harness wars starting, everyone is focused on making agents more capable. I think the harder problem is making their results trustworthy. I quit my job and started a web-services company because, honestly, why not. I help small businesses and freelancers figure out where AI is genuinely useful. My agents perform most of the observable work, while I retain final judgment and approve anything affecting a client or the live business. That led me to build a system for turning my operating judgment into something explicit, testable, and reusable by AI agents. There are about 20 scheduled workers operating across four workspaces and three AI vendors. They share one memory system, but a human must approve anything that affects a client or the live business. For three weeks, I stress-tested the system and documented what worked, what failed, and what only looked convincing at first. The biggest lesson: two AI agents agreeing does not automatically mean the answer is reliable. We had two “independent” reviewers agree on 12 out of 14 decisions. That looked impressive until we realized they were both the same model. It was basically the same brain sitting in two chairs. Adding a different model will make future comparisons meaningful, but it cannot make the old results more trustworthy after the fact. A few other findings: We thought one worker had no access to account credentials. Then it revealed that its session had quietly inherited around 100 connector tools. Our earlier audits missed them because we checked from the operator’s computer, not from inside the worker’s actual environment. A carefully selected 13 KB set of operating principles beat a 20 KB package containing all the directly relevant source material. The larger package even contained the exact rule needed to avoid the mistake—and still made it twice. Giving an AI the right information does not mean it will apply it. Two tightly controlled experiments produced no measurable improvement. We shipped nothing from them, but included the failures in the report. A record that hides its misses cannot be trusted when it claims a win. GrokBot now helps run our order desk. It cannot directly change anything; it can only prepare a proposal for human approval. Its limits are enforced by the database itself, not merely written in a prompt. During its first shift, it handled 21 calls without attempting anything outside its lane. I turned the results into three papers: The case study shows what happened. The technical report explains how to rebuild and test the system. The white paper explains the larger idea behind it. Link in the comments below
The article discusses the criteria for trusting AI agents with real-world tasks, questioning the balance between usefulness and risk, and seeks insights from users on practical workflows.
The author observes that AI agents exhibit human-like failure patterns, such as overconfidence and skipping steps under context pressure, suggesting that system reliability depends more on robust validation and controlled environments than just model intelligence.
The article discusses the key factors, such as reliability, error handling, and transparency, needed to trust AI agents for real business work, beyond just model intelligence.
The author conducted a test comparing multiple AI models on a research task and found that models sometimes confidently disagree. They suggest that AI agents should consider multiple model opinions for important decisions like planning, code review, or research, and ask how others handle this.
An exploration of how using multiple AI models for agent workflows reveals hidden uncertainties and reasoning gaps, suggesting that future systems may rely on cross-model consensus rather than single-model chains.