Using Claude as a pair-programmer on a real MuleSoft project

We spent three months treating a model as a junior on a real MuleSoft project. This is what worked, what stayed firmly with humans, and where the savings actually landed.
Good at the tedious, supervised on the subtle
DataWeave transforms, boilerplate flows and test scaffolding came back fast and mostly correct. Anything touching error handling, transactions or security needed a human in the loop, and we kept it that way on purpose.
The savings were real but unglamorous: less time on the repetitive 60 percent, so the team spent more of its attention on the 40 percent that needed judgement.
More from the FlorenceNext team
Tech10 Feb 2026FlorenceNext TeamCost-aware MuleSoft on AWS: the bits we wish someone had told us
Patterns that bring runtime fabric cost down without sacrificing the latency budget your SLAs depend on.
General27 Jan 2026FlorenceNext TeamThe engineering practices that survive a fast-growing consultancy
A short list of rituals that scale, and the ones we deliberately stopped doing as the team passed 30 people.
Tech13 Jan 2026FlorenceNext TeamFrom SOAP to GraphQL: migration tips from real client projects
Practical notes on schema design, auth, caching, and how to move clients across without a hard cutover.