A recap of what caught our attention at Kaiku this week.
What the data says about AI in software development
DX’s shared a talk on the state of AI in software development draws on data from around 200,000 engineers. In short, deployment frequency is rising, but change failure rate has become far more volatile. Similar to our experience at Kaiku, faster code generation does not mitigate poor fundamentals software development - robust deployment pipelines help keep the increased throughput under control.
Two more details stood out: Junior engineers use AI the most, yet staff-plus engineers save as much time while using fewer tokens. The case studies shared from Morgan Stanley to Zapier and Spotify both apply AI across the whole delivery lifecycle rather than just the coding step.
Fundamentals still matter, for agents too
A post on token consumption describes Thoughtworks’ test of refactoring a monolith into smaller domains in a 150,000-line app built by coding agents. Refactoring one 17,000-line file cut the input tokens for the same change by about 83%. In other words, a codebase that is easier to work in is cheaper to work in, whether the reader is a human or an agent (so not surprises there!).
The obvious caveat is that the experiment started from an app the agents had built themselves. Looking to reality, the harder question is how to break down a genuine legacy monolith, with hundreds of thousands of lines across several old languages and all the business context that is needed within that.
Benchmarking tech stacks for agents
Stack Cap Bench is an early benchmark that measures the token cost and performance of different technology stacks, so far on small projects. Its own advice is to treat the numbers as a direction and not a verdict. However, this is really cool to see and an interesting approach regardless.
It raises the question of whether tech choices will shift towards whatever agents handle best. We think any such shift still has to fit the skills of the people involved as even now someone needs to step in when things go wrong and keep the system architecture sane.
A governed chain from conversation to code
The ANSE project treats requirements, architectural decisions, acceptance criteria, test plans and generated code as artefacts in one governed chain. The output of each stage becomes the controlled input to the next, with a human approving wherever approval changes the outcome. Its flow runs from conversation to requirement, architecture, contract and finally code.
This caught our eye as it closely matches how we have been describing AI-native engineering: the aim is not a better coding assistant but the discipline that makes one accountable.
Want to talk about where AI-native delivery is heading? Get in touch.
