WHAT NEW IN X11.04 WED

AI collections from X · 22 collections · 67,905 posts

gpt-6VIDEO

An OpenAI developer has Astra write the exam before the agent

Instead of having Astra build agents directly, a developer first has it write a strict test (EXAMINER) that an agent must pass to ship; last month 19 of 50 agents failed and were not released.

Codio
@0xCodioArchived 2026.09.20Originally posted 2026.09.19
Codio @0xCodioOriginally posted 2026.09.19
Open original on X

OpenAI developer stopped asking GPT-6 Astra to build agents. He asks it to build the test first - and the agent only exists if it passes. He didn't generate the worker. An agent built to do a job looks done the moment it runs - and you find out it's broken in production, on a real task. EXAMINER is what Astra writes first. Before any worker exists, it builds the hard test that worker will have to survive - the edge cases, the traps, the ways it could fail. BUILDER only then writes the agent. But it's writing toward a known bar - not "do the task," but "pass this exam that already exists." TRIAL runs the new agent against the examiner before you ever see it. Fail, and the agent is thrown out - rewritten, retested, silently, until it clears. BIRTH releases only the survivor. What reaches you isn't a fresh guess - it's an agent that already passed its own hardest test. Here's the part nobody copies. Everyone builds the agent, then tests it. He builds the exam before the agent, because a worker written first gets tested against a bar shaped to flatter it - the test bends to the code instead of the code to the test. Last month Astra spun up 50 agents, and 19 failed their own examiner and never shipped. A build-first setup ships all 50, and finds the 19 the expensive way. Your last broken agent didn't look broken. It ran fine on the cases you thought of - and died on the one you didn't, because the test came after the code. You know this from any real hire. You don't build the role around whoever showed up. You set the bar first, then find who clears it - or you're just justifying the person you already picked. The agent built before its test isn't proven. It's assumed - passing a bar drawn to match it, blind to the failure it was never checked against. Everyone is building agents, then testing them. He makes Astra write the exam first, and only births the agent that already beat it. Don't let this rot in your bookmarks. The article below is the whole system - the examiner Astra writes first, the builder aimed at a known bar, the trial that runs before you see it, and the birth of only the agent that passed.

This post quotes another post: x.com/0xcodio/status/2101235033914724496

x.com/0xcodio/status/2101281362493755759Archived 2026.09.20 10:10