Inspiration Get Shit Done worked but had problems: projects took too long, wasted tokens on simple tasks, and autopilot failed. Users wanted one prompt that just works. I built a system that takes a one-shot prompt, estimates effort, negotiates its own timeline, and delivers.
What I Built Three core agents: Manager – Estimates scope, delegates, negotiates timelines Architect – Designs from UX backward to features to code Engineer – Handles implementation, grounds the Architect's vision Subagents for planning, execution, and verification.
Workflow: Manager → Architect → Engineer → Subagents → product.
How I Built It Every agent must justify its existence or get cut Architect starts from user experience, not backend Manager negotiates between Architect and Engineer to compress timelines Subagents run in isolated contexts to prevent bloat
Challenges: Overcomplication: System over-engineered small projects. Added scope classifier—small projects skip planning. Cut token waste 40%. Architect-Engineer deadlock: Too much arguing. Manager breaks ties: prioritize UX unless implementation cost is too high. Hallucinated libraries: Executor invented dependencies. Verifier now checks against package database. Cut runtime errors 80%. Context bleed: History polluted new projects. Reset context per project; only outputs carry forward. Cut tokens another 20%.
What I Learned The harness matters more than the model. Negotiation beats top-down commands. Minimalism is harder than adding features. Users want results, not agent management. Goal is collaboration, not replacement.
Built With
- pi
Log in or sign up for Devpost to join the conversation.