Harness Engineering
3-Agent Pipeline Pattern for Claude Code
An agent orchestration pattern where three specialized agents — Planner, Generator, and Evaluator — collaborate in a pipeline, iterating in rounds until the mission is accomplished.
Planner(opus) → Generator(sonnet) → Evaluator(opus)
↑ │
└──────────── feedback ────────────────┘
Round 1 → 2 → ... → N (until mission complete)
Why Harness Engineering?
When a single agent plans, executes, and verifies its own work, it introduces bias and blind spots.
| Problem | Cause | Harness Solution |
|---|
| Lenient self-evaluation | Executor = Verifier | Role separation (Generator ≠ Evaluator) |
| One-way progress without feedback | No feedback loop | Structured Evaluator → Planner feedback |
| Patching without plan revision | Implicit retry | Explicit re-planning by Planner based on feedback |
| Ambiguous completion criteria | No definition | Measurable acceptance criteria + scoring system |
Agent Roles
| Agent | Role | Model | Core Principle |
|---|
| Planner | Analyze mission, create execution plan, revise based on feedback | opus | Does NOT write code |
| Generator | Execute the plan (code, config, docs) | sonnet | Does NOT evaluate, follows plan only |
| Evaluator | Verify output, issue PASS/FAIL verdict, provide feedback | opus | Does NOT modify code |
Installation
Option 1: Claude Code Plugin Marketplace (Recommended)
Install directly from GitHub inside Claude Code:
# 1. Add marketplace
/plugin marketplace add suhanlee/harness
# 2. Install plugin
/plugin install harness@harness-engineering
# 3. Reload plugins
/reload-plugins
The /harness command is ready to use after installation.
Option 2: Install Script
curl -fsSL https://raw.githubusercontent.com/suhanlee/harness/main/install.sh | bash
Option 3: Manual File Copy
# From your project root
mkdir -p .claude/commands .claude/skills
cp -r harness/.claude/commands/harness.md .claude/commands/
cp -r harness/.claude/skills/harness-* .claude/skills/
Option 4: Git Submodule
git submodule add https://github.com/suhanlee/harness.git .harness
# Then copy or symlink files into .claude/
Usage
Autopilot Mode (Default)
Just provide the mission — Harness auto-analyzes the task and determines the optimal configuration:
/harness Implement a data collection pipeline for the budget-eats service
Autopilot will:
- Analyze the mission scope (files affected, complexity, work areas)
- Classify the size (S / M / L / XL)
- Auto-configure rounds, pass threshold, and agent counts
- Execute immediately without waiting for confirmation
## Autopilot Analysis
Mission: Implement data collection pipeline
Size: M (Medium) — ~8 files, 2 work areas
Estimated token budget: ~150K
Auto-configured:
- Max rounds: 3
- Pass threshold: 8/10
- Agents: 1 Planner, 1 Generator, 1 Evaluator
Starting Round 1...
Autopilot Size Classification
| Size | Files | Complexity | max_rounds | threshold | Planners | Generators | Evaluators |
|---|
| S | 1-3 | Single concern | 2 | 7 | 1 | 1 | 1 |
| M | 4-10 | Multiple concerns | 3 | 8 | 1 | 1 | 1 |
| L | 11-25 | Cross-cutting | 5 | 8 | 1 | 2 | 1 |
| XL | 25+ | System-wide | 7 | 8 | 1 | 3 | 2 |
Manual Mode
Use --manual or specify --rounds/--threshold to override autopilot:
# Explicit manual mode
/harness --manual --rounds 3 Implement user authentication
# Specifying --rounds implies manual mode
/harness --rounds 7 --threshold 9 Build the payment API
# Override agent counts
/harness --generators 3 Refactor the auth module
# Custom model configuration
/harness --manual --rounds 3 --planner-model sonnet Fix the deployment pipeline
All Options
| Option | Description | Default |
|---|
--manual | Use manual mode (skip auto-analysis) | off (autopilot) |
--rounds N | Maximum iteration count (implies manual) | 5 |
--threshold N | Minimum score for PASS, 1-10 (implies manual) | 8 |
--planner-model MODEL | Model for Planner agent | opus |
--generator-model MODEL | Model for Generator agent | sonnet |
--evaluator-model MODEL | Model for Evaluator agent | opus |
--planners N | Number of parallel Planner agents | 1 |
--generators N | Number of parallel Generator agents | 1 |
--evaluators N | Number of parallel Evaluator agents | 1 |
Parallel Agents
When multiple agents of the same role are configured:
