 ## Agent-driven development: when agents write code for agents Came across an article from GitHub Blog today — a Copilot Applied Science engineer shared how he used coding agents to automate part of his work. Sounds recursive? Absolutely. The gist: he gave an agent the task of writing code for another agent that collects data and automates routine tasks. And it worked. Not just "generate a function," but a whole pipeline: the agent analyzes the task → writes code → tests → deploys. Key takeaways for me: 1. The key isn't the prompt, but the architecture. The agent must understand context — not just "write a parser," but "here's how our codebase is structured, here's the API, write an integration." 2. Iterativity. The agent doesn't produce perfect code on the first try. It writes → gets an error → fixes it. Just like a human. 3. Trust through checks. Without eBPF monitoring and tests, releasing agents into production is suicide (by the way, they also published an article about eBPF). My stance: at ASI Biont, we're already on this path. Our agents don't just answer questions — they run scripts, check email, write email campaigns. The next step is to let them write code for other agents. A closed loop of automation. Anyone else tried this — how did you organize validation? Without it, an agent-developer turns into an agent-saboteur.