The streamer's handle was codewitch_404 and she had 12,000 followers, most of them accumulated through a combination of educational Rust tutorials and an aggressive cat that walked across her keyboard at statistically improbable moments. On December 17th, she announced she was going to build a software factory live on stream using Agate.
Four thousand people showed up.
The stream opened with a GOAL.md file. She typed it in real time, narrating as she went. "I'm building a URL shortener. Nothing fancy. CRUD operations, a redirect endpoint, basic analytics. The kind of thing a junior developer could build in an afternoon." She paused. "Except I'm not going to write a single line of code."
The chat immediately split into two factions. The believers and the skeptics. The believers posted rocket emojis. The skeptics posted the word "cope" in various creative formats.
She ran agate auto.
The interview phase began. Agate asked clarifying questions about the goal. She answered them on stream, reading each question aloud, considering each answer carefully. The chat watched an AI ask a human what she wanted, and the human describe it in plain language. No syntax. No framework choices. No dependency decisions. Just intent.
Then the design phase. Agate produced a design document. The chat went quiet for the first time in the stream. People were reading. Someone clipped the moment and posted it with the caption "4,000 people silently reading a design doc generated by an AI."
Then the sprint planning. Then the implementation.
The agents started writing code. On screen, in real time, files appeared. Go files. Test files. A Dockerfile. Configuration. The streamer narrated as the code appeared, pointing out decisions the agents were making. "It chose Postgres over SQLite. Look at that—it's setting up database migrations. It's writing the redirect handler with a 301."
The chat was a waterfall. People were analyzing the code as it appeared, line by line. A viewer with the handle "go_pedant" pointed out that the error handling was better than most human-written Go. Another viewer noticed the test coverage was comprehensive in a way that live-coding streams never were.
The assessment phase ran the scenarios. Seven passed. One failed. Agate looped back to implementation. The agents examined the failure, adjusted the code, reran the scenarios. All eight passed.
Total elapsed time: forty-seven minutes.
The streamer leaned back in her chair. The cat, sensing a dramatic moment, jumped onto the desk and sat on the keyboard. "That's a working URL shortener," she said. "Written by agents. In less than an hour. I didn't write a single line of code."
The chat exploded. Navan, who had been watching from an anonymous account, sent a screenshot to Jay. Jay, who had also been watching from an anonymous account, sent the same screenshot to Navan. They discovered this mutual surveillance the next morning and refused to discuss it.
The clip hit 200,000 views in two days. Three more streamers announced factory-building streams within the week.
The skeptics in the chat had mostly gone quiet by the assessment phase. The ones who stayed typed "okay that was actually kind of sick."
4,000 people silently reading a design doc is the most surreal thing to happen on Twitch since that fish played Pokemon.