I have been experimenting a bit with orchestrators that start with a spec and operate fully autonomously from that point on. In the interest of taking it too far I took a stab at automating the problem discovery process itself. I built a quick prototype of PAT - my “problem acquisition tool”, and let it run overnight for n=3 projects. It scanned RSS feeds, ranked projects, then sent them to project orchestrators to be built.
A few of the key design decisions for the orchestrator process:
Pi-Agent with Claude Opus
- Pi-agent (a minimal coding agent harness) provides full control of the context and process without model lock-in.
File system as the db
- Leans heavily on Markdown with frontmatter.
- An Analyze bash tool uses duckdb and the yaml extension for comprehensive project visibility with SQL over files
Fully headless project orchestrators
- Only project interface is an inbox that takes markdown tasks.
- All events and llm responses logged
”Portfolio Manger” coordinator
- Specs out new projects
- Interfaces with headless project orchestrators
- Watches development and gives status notifications via tools
Heavily sandboxed
- Separate accounts for everything and running on a cloud vm
So here were the results - not exactly the best ideas yet! But certainly an interesting outcome:
