AI Code Review and the Age of Autonomy with James Governor

AI Code Review and the Age of Autonomy with James Governor

Share via Twitter Share via Facebook Share via Linkedin Share via Reddit

Link to podcast

In 2026, banks started telling James Governor they would no longer have humans review all of their code. AI code review went from experiment to default faster than almost anyone predicted, including the analyst firm that has spent two decades calling where developers go next. I’m joined by James, co-founder of RedMonk and organizer of Monki Gras in London, to trace how developers stopped asking for permission and what happens now that agents are doing the same.

RedMonk started around 2002 as the open source analyst company, at a time when enterprises claimed they used no open source and getting an application into production could mean waiting six months for hardware. James frames everything since as a story about permission. Open source removed the need to ask before using software, and AWS removed it for infrastructure. Its first product was a simple queuing API, not storage or compute. Running YouTube on Oracle, James reckons, would have cost more than the world’s GDP.

We also revisit Heroku, which James calls truly magical, and he sees Vercel as its closest successor. I moved to London around 2012 and landed in the middle of the Kubernetes scene, so I ask why the city punched so far above its weight. James points to the BBC as a training ground for building at scale, cheap Shoreditch rents, government-backed inward investment, and a few investors who meant founders no longer had to leave.

Then we get to AI. James calls 2025 the year of agents and 2026 the year of autonomy. Model quality jumped around November 2025, and by spring 2026 even banks were telling him they would stop reading all of the code. My concern is liability. Frameworks like the Cyber Resilience Act require code review, but they do not say a human has to do it, and when something breaks, someone still has to answer for it. James expects most of that review to be done by other agents, because firms competing with Revolut and Monzo cannot afford to sit it out.

On open source, James argues the hard part was never building the thing but running it, and that maintainers were struggling long before the flood of agent pull requests. He knows at least two people building their own databases and warns that snowflakes are beautiful until they melt. I’m on the fence about the SaaSpocalypse: bloated SaaS will feel it, but I expect many niche products that were not viable before.

Daniel Stenberg of curl finds current models very good at finding problems and less good at fixing them. James adds that they are even better at exploiting them, and shares Matthew Prince’s claim that many security vendors will disappear within two to three years as models write safer code. I think it is closer to ten, given how much existing code is out there.

We close on the bubble. RedMonk was founded coming out of the dot-com crash, and James would be surprised to get through 2027 without real carnage. But the shakeout is the beginning, not the end. He expects Anthropic to still be around in ten years. My long bet is Apple, which has the cash to wait it out and the hardware to run models locally.

For founders and engineering teams, the takeaway from this episode is practical: decide who is accountable for code no human has read, and build AI code review and audit trails around that answer. Count the cost of running something for years before building your own version. Plan for more vulnerability reports in the short term, and treat AI as a workflow change you start now rather than a vendor choice you make later.