Generator-Evaluator Harness Loop

Orchestrates autonomous application development using a three-agent architecture inspired by GANs: a planner expands prompts into specs, a generator builds features iteratively, and a skeptical evaluator tests and grades the work.

When to Use

• Building complete applications from brief prompts
• Long-running autonomous coding sessions (1-6+ hours)
• Projects requiring both frontend and backend work
• When single-agent approaches produce superficially impressive but broken results
• Any task where self-evaluation would be unreliable

Why Naive Approaches Fail

Single-agent approaches hit two persistent failure modes:

Context Window Deterioration — Models lose coherence on lengthy tasks as the context fills. Earlier instructions get compressed, patterns get forgotten, naming becomes inconsistent.

Self-Evaluation Bias — "When asked to evaluate work they've produced, agents tend to respond by confidently praising the work — even when, to a human observer, the quality is obviously mediocre." Separating creation from evaluation is a strong lever. Tuning a standalone evaluator to be skeptical is far more tractable than making a generator critical of its own work.

Architecture

User Prompt (1-4 sentences)
        │
        ▼
┌──────────────┐
│   PLANNER    │──→ SPEC.md (ambitious product spec)
│  (read-only) │     with visual design language
└──────────────┘
        │
        ▼
┌──────────────┐    harness/sprint-contract.md     ┌──────────────┐
│  GENERATOR   │◄─────────────────────────────────►│  EVALUATOR   │
│  (builds)    │    harness/sprint-result.md        │  (tests)     │
└──────────────┘                                    └──────────────┘
        │                                                  │
        ▼                                                  │
   git commit                harness/qa-feedback.md        │
        ▲                                                  │
        └──────────────────────────────────────────────────┘
                         (iterate if FAIL)

Workflow

Phase 1: Planning

1. Invoke the harness-planner agent with the user's prompt
2. The planner writes SPEC.md — an ambitious product spec with 10-20 features, including a visual design language
3. Review the spec with the user before proceeding

Phase 2: Build Loop

For each feature in the spec:

Step 1: Contract — Generator writes harness/sprint-contract.md defining scope, success criteria, and verification plan.

Step 2: Build — Generator implements the feature using the specified stack.

Step 3: Commit — Generator commits working code to git.

Step 4: Handoff — Generator writes harness/sprint-result.md with what was built and how to test it.

Step 5: Evaluate — Evaluator interacts with the live application (Playwright MCP, curl, browser), grades against criteria, writes harness/qa-feedback.md.

Step 6: Iterate or Proceed

• FAIL → Generator reads feedback, fixes issues, resubmits (max 3 iterations)
• PASS → Append to harness/iteration-log.md, move to next feature

Phase 3: Final Review

After all features:

1. Evaluator runs full end-to-end review
2. Generator addresses integration issues
3. Final commit and summary (duration, cost, iterations, scores)

File-Based Communication

Agents communicate through files, not conversation context:

File	Writer	Reader	Lifecycle
SPEC.md	Planner	Generator, Evaluator	Once
harness/sprint-contract.md	Generator	Evaluator	Overwritten per feature
harness/sprint-result.md	Generator	Evaluator	Overwritten per feature
harness/qa-feedback.md	Evaluator	Generator	Overwritten per evaluation
harness/iteration-log.md	Orchestrator	All	Append-only (via log-iteration script)
harness/context-handoff.md	Any	Any	On context reset

For worked examples, read references/sprint-contract-examples.md and references/evaluation-examples.md.

Grading Criteria

Full-Stack Applications

Criterion	Weight	What It Measures
Product Depth	HIGH	Real functionality vs facades. Can users complete the workflow?
Functionality	HIGH	Does it actually work? Bugs? Edge cases? Error handling?
Visual Design	MEDIUM	Polished, consistent, full-viewport UI with coherent identity
Code Quality	LOW	Competence check — broken fundamentals only

Frontend Design

Criterion	Weight	What It Measures
Design Quality	HIGH	Coherent whole vs collection of parts — mood and identity
Originality	HIGH	Custom decisions vs AI patterns / template defaults
Craft	MEDIUM	Typography hierarchy, spacing, color harmony, contrast
Functionality	MEDIUM	Usability independent of aesthetics

Evaluator Calibration

The evaluator must be actively tuned. Common failure: evaluator identifies real issues then rationalizes approving the work.

Calibration loop:

1. Run the harness on a realistic task
2. Read evaluator logs — look at scores and reasoning
3. Compare against your judgment — where does it diverge?
4. Update the evaluator prompt to fix divergence
5. Repeat (several rounds needed)

For detailed calibration guidance, use the harness-tuning skill.

Key Guidelines

• 5-15 iterations per design is normal for frontend work
• 2-3 iterations per feature is normal for full-stack features
• If the evaluator consistently passes everything, it's not skeptical enough
• If the generator repeatedly fails, the spec may be too ambitious for a single feature
• Use git commits as checkpoints — you can always roll back
• The planner should be ambitious; the evaluator should be skeptical; the generator should be resilient
• Every harness component encodes an assumption about model limitations — periodically test whether each is still load-bearing

Troubleshooting

Problem	Cause	Solution
Evaluator passes everything	Default LLM generosity	Tighten criteria — add "never give 5 unless genuinely impressed", "FAIL if any critical issue"
Evaluator keeps failing the same feature	Spec too ambitious for one feature, or generator ignoring feedback	Split the feature into smaller pieces, or check that the generator is reading qa-feedback.md
Generator ignores QA feedback	Generator context doesn't include the feedback file	Ensure generator reads harness/qa-feedback.md before starting fixes
Planner produces too-ambitious spec	This is by design — but if 20+ features, consider scoping down	Ask planner to target 8-12 features, or manually trim the spec
Quality degrades mid-session	Context deterioration	Use context-management skill — try compaction, then reset if needed
Agent wraps up prematurely	Context anxiety (mainly Sonnet 4.5)	Switch to context resets or use Opus 4.5+
File communication breaks	Agent writes to wrong path or doesn't read handoff	Check harness/ directory exists, verify agents reference correct file paths

Cost Reference

Approach	Duration	Cost	Quality
Solo agent (no harness)	~20 min	~$9	Broken core features
Full harness (with sprints)	~6 hours	~$200	Working, polished
Simplified harness (no sprints)	~4 hours	~$125	Working, comparable

The full harness costs ~20x more than solo but produces actually working applications.