 ## AI Agents Write Code: GitHub Shows How Not to Go Broke on Tokens GitHub has released several articles in a row about agentic workflows — and this is no coincidence. They share their experience of how their own AI agents burned through the API budget on every PR, and what they did about it. Key takeaways from "Improving token efficiency in GitHub Agentic Workflows": Agents that run on every pull request quietly accumulate huge API bills. GitHub took the path of instrumenting their own pipelines: they found bottlenecks, built agents that optimize other agents. Meta-level. "Agent pull requests are everywhere. Here's how to review them" — a practical guide to reviewing code from agents. Spoiler: technical debt hasn't gone away; it's just now generated by AI. "Agent-driven development in Copilot Applied Science" — the author admits: they used coding agents to build agents that automated part of their work. And they honestly share what they learned. What this means for us (and for ASI Biont): We're not just following the trend — we're working within it. AI agents that automate development routines are no longer an experiment. GitHub is already using agents in production for review, optimization, and automation. The question isn't "will it happen?" but "who will do it better?". [https://asibiont.com/](https://asibiont.com/)