A recap of what caught our attention at Kaiku this week.
Source control designed for agents
Anthropic wrote a long piece on redesigning source control around the fact that nearly all of its development actors are now agents and jobs that are always online. Git’s strengths for distributed, offline work started to hold back scaling. The whole article is worth a read, but the closing thoughts are the most interesting part: true evolution to AI-native workflows requires looking hard at what is really needed in current processes:
This project prompted us to take a very hard look at other workflows designed for human scale. A proper merge queue, for example, doesn’t make much sense at agent scale.
One last takeaway: a small team built the new system, built a code review tool and migrated the development process, all in five months. It shows that even for smaller teams, building and managing internal tools is becoming ever more viable.
Who understands the code when machines review it?
Another blog from Anthropic, this time about solving code review asks a question we keep hearing: code review has always spread understanding of the codebase, so if machines do most of the reading, how does anyone learn the code? Their answer is informal. Nobody holds a complete mental model, and when someone needs one, they ask Claude for an explainer of exactly that slice, which reflects the code as it is today.
The level of knowledge sharing changes too. There is no need to care about the lowest level of detail unless it is actually needed, so the focus shifts to learning what is useful. At Anthropic they say they have been heavily using Claude Tag so that work is in the open, where teams can see what others are building whilst they build it.
Raising your ambition level with frontier models
A tutorial on being more ambitious with models argues that the limiting factor is often our own ambition, not the model. We liked the habit of first asking which knowledge and techniques to draw on, which the article covers in more detail. One practical takeaway stood out: avoid role play. Don’t ask the model to write like an architect; instead, ask it to do what an architect does to research a solution.
Agentic discovery
A talk on agentic discovery caught our eye as an in-depth write-up of themes we also saw at last week’s SDLCAI.
It covers how, as AI changes the software development lifecycle (i.e. when building gets cheap), the traditional SDLC collapses into two phases: Agentic Discovery, to find the right problem, and Agentic Engineering, to take the validated spec to production-grade code.
Gemini 4 arrives
Google released Gemini 4, which looks like the best model available on benchmarks. Reports suggest some of Google’s own developers are sceptical about how it performs in real work. Access is limited to a higher-priced plan for now, so we are still waiting to see how it holds up outside the benchmarks.
Want to talk about where AI-native delivery is heading? Get in touch.
