The core design decision was bounded responsibility per agent rather than one monolithic prompt. 38 agents in 5 categories (governance, knowledge, technical, business, evolution), each with a defined role and explicit collaboration interfaces. Workflows are pipelines with ordered handoffs, not free-form agent swarms. A technical task goes Architect → Security-Expert → Engineer → Security-Expert (second pass) → QA → Reviewer → Documenter → Historian. The routing is deterministic.
State is plain markdown in an `active/` directory that loads on session start. Git handles persistence and backup. Two-remote model separates the public framework from private user data so you can receive updates without exposing your files.
The part I find most interesting architecturally: the evolution loop. An evaluator agent measures outcomes, an improver proposes changes through a formal queue, and nothing gets implemented without user approval. The system's structure is mutable by design but gated on human review.
Tradeoffs worth noting: it's opinionated. The agent definitions and workflow ordering reflect how I work as a solo founder. If your workflow is fundamentally different you'll want to modify the agent prompts and pipeline ordering. The `/setup` wizard handles initial customization but it's not infinitely flexible out of the box.
MIT licensed. Happy to discuss architecture decisions.