Q&A with Simon Willison on the November release of GPT-5.1 and Opus 4.5 as the inflection point for coding, exhaustion due to managing coding agents, and more
Simon Willison is a prolific independent software developer, a blogger, and one of the most visible and trusted voices on the impact AI is having on builders.
Lenny's NewsletterLenny Rachitsky
Context & Ripple Effects
The discussion lands after a mixed progression in coding-model claims: GPT-4.5 was reported to be competitive on coding benchmarks, while later reporting pointed to GPT-5 gains in practical software engineering. Willison’s framing puts the emphasis not on a single benchmark win but on whether models can reliably carry more of the development workflow.
That distinction matters because OpenAI’s subsequent Codex work on long-horizon changes and context compaction points toward agents operating across larger codebases. The limiting factor may increasingly be the developer’s ability to delegate, inspect, and coordinate agents rather than prompt a single model.
First-order effects
Developers evaluating GPT-5.1 and Opus 4.5 will test them for sustained coding work and agent supervision, not merely isolated code generation; Willison’s reported exhaustion makes the coordination burden a near-term adoption constraint.
The conversation raises the practical bar for coding assistants: capability gains matter only if they reduce the human overhead of managing multiple agent runs and reviewing their output.
Teams adopting coding agents are likely to invest more in review checkpoints and clearer task decomposition, because greater agent autonomy can shift work from writing code to validating and directing it.
Third-order effects
If coding agents keep improving on longer tasks, software development could reorganize around smaller groups directing agentic production, with human judgment concentrated in specification, integration, and assurance.
The reported management fatigue suggests a countervailing structural need for operational controls and better agent interfaces; without them, capability improvements may not translate proportionally into deployment.
The trend: Coding AI is moving from autocomplete-style assistance toward agent-managed software production, where orchestration and assurance become as important as model capability.
“Using coding agents well is taking every inch of my 25 years of experience as a software engineer, and it is mentally exhausting. I can fire up four agents in parallel and have them work on four different problems, and by 11am I am wiped out for the day. There is a limit on [vid…
The best analogy of how I use coding agents is pair programming. They write the code, I review it. They write the spec, I think about the next problem (sometimes using agents too!). Etc. The only time I ever run more agents (3+) is when I absolutely don't care about the code
Tip: Figure out your personal ceiling for running multiple agents in parallel. We need to accept that more agents running doesn't mean more of _you_ available. The narrative is still mostly about throughput and parallelism, but almost nobody's talking about what it actually
This is very true in my experience Programming used to feel kind of leisurely and relaxing, like a crossword or sudoku Now it feels more like debate. My raw horsepower is at the limit of my ability to absorb complex information and make decisions with real consequences.
I was a guest on @lennysan's podcast! We talked about agentic engineering and all sorts of other LLM-related topics for 1h39m(!), plus a little bit about kākāpō parrots - here's my selection of highlights from our conversation https://simonwillison.net/...
it is great to see some sanity around the fact that agents increase the need for solid computer science fundamentals and software engineering skill, and using them well takes effort as opposed to the frenetic and hypoxic claims the machines will replace all coders
simon's spot on here and i'm glad someone so prominent is stating this publicly. on top of this, coding agents and the human reward system interact in novel and strange ways we'll be unpacking for years to come imo so stay safe out there, people!
I can relate so so much to this. I would also stress, let's not forget about “enjoying” the process of building, and not just the destination. These tools bring us so much closer to the destination, that we might fall into the trap of forgetting about the journey.
🚨 Most of the AI industry seems to ignore what @simonw is saying here. There is a biological LIMIT to humans' cognitive load, and pushing beyond it will likely lead to lower-quality decision-making and ethical/legal risks. I call it the acceleration paradox (more info below).
My biggest takeaways from @simonw: 1. November 2025 was an inflection point for AI coding. GPT 5.1 and Claude Opus 4.5 crossed a threshold where coding agents went from “mostly works” to “almost always does what you want it to do.” Software engineers who tinkered over the
I asked @simonw what the next leap in AI software engineering is likely to be. He explained the “dark factory” pattern where teams don't write any code or even look at their code. [video]
“Using coding agents well is taking every inch of my 25 years of experience as a software engineer.” Simon Willison (@simonw) is one of the most prolific independent software engineers and most trusted voices on how AI is changing the craft of building software. He co-created