Meta Wants Machine‑Readable Code

Meta is telling engineers to change code that already works so software bots can read it like a street map instead of a maze. The shift is not about prettier code comments; it is about making artificial intelligence agents able to find files, follow rules, and change the right thing without wandering for hours. (timesofindia.indiatimes.com) That only makes sense if you know how code assistants fail inside big companies. A model can write a neat function in isolation, but a production codebase can span thousands of files, several programming languages, and years of unwritten team habits. (engineering.fb.com) Meta described that problem in one of its own engineering posts on April 6, 2026. In a single data pipeline, its agents had to work across four repositories, three languages, and more than 4,100 files, and the agents were not making useful edits fast enough. (engineering.fb.com) So Meta built what is basically a guidebook for the bots before letting them code. The company used more than 50 specialized agents to read every file and produce 59 context documents that captured design choices and “tribal knowledge” that human engineers usually carry in their heads. (engineering.fb.com) After that mapping work, Meta said its agents had structured navigation guides for 100% of the modules in that pipeline, up from 5% before. Preliminary tests showed 40% fewer tool calls per task, which means the agents spent less time poking around and more time editing the right code. (engineering.fb.com) That is the backdrop for Zuckerberg’s new push. If the company wants agents to write and modify real production software, the codebase has to expose clear entry points, stable naming, and obvious relationships in the same way a well-labeled warehouse lets a robot find the right shelf. (timesofindia.indiatimes.com) Meta is not treating this as a side experiment. Internal targets reported in late March 2026 said some teams were expected to have more than 75% of committed code written with artificial intelligence assistance, while the Scalable Machine Learning group had targets ranging from 50% to 80% artificial intelligence assisted code. (timesofindia.indiatimes.com) The company is also measuring usage at a scale that sounds more like cloud accounting than software coaching. The Information details relayed by Times of India said Meta logged more than 60 trillion tokens of employee artificial intelligence use over a recent 30-day period and tracked that activity on internal dashboards. (timesofindia.indiatimes.com) This changes what “good code” means inside a company like Meta. For years, engineers optimized for human teammates reading the code later; now they are also being asked to optimize for machine teammates that need explicit structure because they cannot infer unwritten rules reliably. (timesofindia.indiatimes.com)