A factory that builds agent factories for Claude Code

What it does

Harness is a Claude Code plugin that generates agent teams. You say “build a harness for this project” and it spits out agent definitions in .claude/agents/ and skills in .claude/skills/ tailored to whatever domain you described. The generated teams follow one of six preset architectural patterns — pipeline, fan-out/fan-in, expert pool, producer-reviewer, supervisor, or hierarchical delegation.

The interesting bit

The project bills itself as an “L3 Meta-Factory” — a factory that makes other factories, not a factory that makes products. That sounds like taxonomy theater, but the actual trick is decent: it encodes six team topologies as reusable templates, so you aren’t reinventing how agents hand off work every time. The README also includes a controlled A/B test claiming +60% quality scores and 100% win rate across 15 software engineering tasks when using structured harnesses versus raw prompting — though the truncated source cuts off the methodology details.

Key highlights

• Six predefined team architecture patterns with orchestration templates and inter-agent data passing
• Auto-generates skills using “Progressive Disclosure” for context management
• Two execution modes: full Agent Teams (multi-agent collaboration) or lightweight Subagents (single tasks)
• Validation phase includes dry-run testing and with-skill vs without-skill comparison tests
• Companion repo harness-100 ships 200 prebuilt packages across 10 domains
• Supports English, Korean, and Japanese prompts out of the box

Caveats

• The A/B test methodology is truncated in the source; claims look promising but can’t be fully verified from the README alone
• Requires Claude Code’s plugin system — not portable to other LLM runtimes (a separate Codex port exists from another author)
• The “L3 Meta-Factory” framing is more ecosystem positioning than technical distinction

Verdict

Worth a look if you’re already living in Claude Code and find yourself repeatedly assembling ad-hoc agent teams. Skip it if you need runtime-agnostic orchestration or deterministic reproducibility — the README explicitly points to Archon for that adjacent layer.