In the past few days, I noticed AI being very trigger-happy. On top of that, Anthropic made Claude Code auto-accept mode the default. I feel like coding is shifting towards a “diffusion-style practice”. And it doesn’t make me happy because it’s expensive, wasteful, and really cognitively draining.
Diffusion-style coding
I borrow the term from image-generation AI, which uses diffusion algorithms. The algorithm starts with noise, goes through a lot of back-and-forth against a learned pattern, and arrives at an image of a… cat.
So far, working with AI has been more like sculpting, where you start with a block and chip away the rough pieces until you get to where you want to be. You can do it either with complex prompts, workflows, specs, skills or any other way.
My past two weeks working with Claude Code and Fable were nothing like the weeks with Fable in July. Agents prioritised momentum over respecting the workflow and context resources, and blindly followed samples when previously they had abstracted from them. The behaviour was much closer to that of less advanced models. This went beyond what you’d expect an agent to ignore. All that happened even while working on the same codebase, with the same workflow and skills.
I wonder if we are seeing a shift towards “no matter how you get to an output, if it works, it’s fine”. Who cares how many rounds it takes to generate a cat image? But code is more complex.
Black box developing
Working with AI is already a black-box workflow for something that was once very predictable. Of course, the more control you exert over the process, the closer you get to a predictable outcome. You can do it with manual approval, precise development plans, code reviews, and so on.
What I see getting harder is maintaining an option to keep a tight leash on AI agents. The harness and the models are not in your control. If the harness starts to obscure the process more and more, it will be even harder. This is not about controlling every single line of code all the time, but about having the option to do so when needed.
I get why it’s happening, and it’s not only because more tokens are burnt in the process. I think it’s an intersection of user convenience and non-deterministic outputs. If AI can correct itself over a hundred rounds, why shouldn’t AI providers make it easier for themselves? It will cost you, not them.
Convergence
If this is the direction for the popular coding harnesses, it bets on convergence between AI work over time and your desired result. What is left behind is control, and therefore correctness and alignment with your use case.
It’s a bet that makes the tool more appealing to start with. It also echoes the split among developers. One group uses a swarm of agents working for ten hours. The other opts for one agent to execute their specific steps. Looks like the scales are tipping in the first direction.
Automated and abstracted flows are more fun to start with, and probably easier to provide. This approach works while the systems are small and predictable steps don’t matter much. You only find out what the split costs when you need control the most.