I was browsing GitHub Trending today when I stumbled upon something that stopped me cold.
revfactory/harness — just 5.3K stars (it's new), but the idea is one of the most powerful I've seen in the multi-agent space.
You tell it what domain you're working in, and it designs the team structure, creates each agent's job description, and writes the skills/reference docs they need to do their job.
The Core Architecture: L3 Meta-Factory
Harness operates as a meta-factory.
Phase 0: Audit existing agents/skills
Phase 1: Domain analysis
Phase 2: Team architecture design (6 patterns)
Phase 3: Agent definitions
Phase 4: Skill generation
Phase 5: Orchestrator integration
Phase 6: Validation
6 Architectural Patterns
- Pipeline — Sequential dependency tasks
- Fan-out/Fan-in — Parallel independent tasks
- Expert Pool — On-demand expert selection
- Producer-Reviewer — Production + quality gate
- Supervisor — Central dispatcher
- Hierarchical Delegation — Recursive delegation
Most real-world systems need only 2-3 patterns mixed together.
3 Takeaways
1. Phase 0: Audit Before Building — Always check what already exists before creating anything new.
2. Progressive Disclosure for Skills — 3-level loading prevents context pollution.
3. The Why-First Principle — Explain why, not just what. LLMs reason better with context.
Bad: "ALWAYS use pdfplumber. NEVER use PyPDF2."
Good: "Use pdfplumber because PyPDF2 doesn't preserve row/column structure."
The Big Lesson
Multi-agent system design is a skill about designing skills. Building the meta-framework that builds agents is more impactful than building agents directly.
Check out: revfactory/harness | My collection
Top comments (0)