# Coding agents make mistakes. So what? > Source: > Published: 2026-08-19 09:00:00+00:00 Regular readers will probably have figured out that I’m generally an optimist. I think things are always getting better, and the future of the human race looks pretty good. Technology has taken us from grinding poverty to a pretty good life here on Planet Earth. I also believe that life has gotten immensely better for software developers just in the last year. We’ve seen AI agents suddenly and stunningly write real, working code from prompts alone. We tell the AI what we want, and it does it. I think it is fantastic. I’ve used coding agents to build complete, working websites in a weekend. But I meet a lot of skeptics. Some see the agents doing their thing and love to point out the flaws, mistakes, and bugs that the agents produce. They love to mention that AI agents are not compilers and that they aren’t deterministic — the same prompt will produce different results every time. Well, okay. But you know what else makes coding mistakes, writes bugs, and never writes the same code twice when prompted? Humans. Us. Yep, we do those very things. All the time. Not only that, but we have a large infrastructure of tooling to account for our flaws. We have source version control so we can unroll our mistakes and misadventures. We have pull requests and code reviews to prevent our flaws from getting into the codebase. We have whole teams of people — quality assurance engineers — who make sure that what we wrote does what it is supposed to do. We built these tools and this structure all for the purpose of catching errors and for keeping control over the errant ways of humans. And yet, somehow, the fact that AI agents make mistakes is supposed to stop us from using them. Huh. Since the beginning of software development, we’ve been dealing with flaws in the whole system. Many flaws. In fact, it’s a rare commit of any substance that doesn’t have flaws in it. There are bugs everywhere. We hunt them down, log them, track them, and fix them. Our fixes introduce new bugs. We are, after all, flawed beings. Neither humans nor coding agents produce deterministic results. But remember, they both produce code — which is deterministic. When humans write code, they read the specification, formulate a plan, and write the code. An agent does the same thing. And in both cases, the code produced is the thing that defines the system. How that code gets produced shouldn’t make a difference. So, I have no idea why anyone would hesitate to add a new “team member” that, while writing flaws, at least writes them 100 times faster than we humans can. And not only that, they can find the flaws 100 times faster as well. Agents are so good at finding flaws that Microsoft has started using Anthropic’s [Claude Mythos model](https://www.csoonline.com/article/4198019/claude-mythos-faq-capabilities-access-competitors-implications.html) to help harden Windows and Azure against vulnerabilities — finding in minutes flaws that human security teams would take weeks to find. If you could find a human that was half as capable at coding or finding bugs as today’s frontier models, you’d hire her in a heartbeat. And with models rapidly improving, not hiring that coding agent is a foolish omission.