Developer uses Astra to build agents that grow from failure
Instead of writing a feature list up front, the developer has Astra start from the smallest possible agent and add only the one capability each time it hits a wall, reportedly growing from one function to nine.
OpenAI developer builds agents with GPT-6 Astra that start too small to work. The agent grows itself, only when it hits a wall it can't clear. He didn't spec the full agent. A feature list written up front is a guess - half the powers go unused, and the one that mattered was never on it. SEED is where Astra starts - the smallest possible agent. One job, no extras, deliberately missing most of what it'll eventually need. RUN lets it fail. The agent hits the task and breaks on the first thing it can't do - and that break is the signal, not a bug. DIAGNOSE reads the failure. Not "add everything" - just the one capability the wall demanded, named precisely by what stopped it. GROW hands that one addition back to Astra, which writes it in. The agent gains exactly what reality asked for, then runs again. Here's the part nobody copies. Everyone specs the whole agent before it runs. He lets it grow from its own failures, because a capability list written up front is built on imagined needs - and the agent ends up bloated with powers it never uses, missing the one it does. Last month an agent started with one function and grew to nine - each earned by a real wall, none guessed. A spec-first build shipped twenty, and used seven. Your last over-built tool didn't fail on quality. It drowned in features nobody needed - every one a guess about a problem that never showed up. You know this from any product. You don't build every feature you imagined. You ship the smallest thing, watch where it breaks, and add only what the breaking demanded. The agent built to spec isn't complete. It's speculative - heavy with imagined needs, light on the one the task actually turned out to require. Everyone is building the full agent from a plan. He makes Astra grow one from its own failures, adding only what a real wall proved it needed. Don't let this rot in your bookmarks. The article below is the whole system - the seed too small to work, the run that lets it break, the diagnose that names the one missing power, and the grow step that adds only what reality demanded.
This post quotes another post: x.com/0xcodio/status/2101235033914724496