Sprint Contract — Multi-Agent Quality System
Based on Anthropic's harness design for long-running apps: separate the agent doing the work from the agent judging it.
Core Principle
Never let the builder evaluate their own work on complex tasks. LLMs reliably praise their own output — even when it's mediocre. An independent evaluator, tuned to be skeptical, catches what self-evaluation misses.
Architecture
Planner (you/human) → Generator (sub-agent) → Evaluator (independent sub-agent)
↑ |
└──────────── feedback loop ←──────────────────┘
Workflow
1. Write a BRIEF.md with Sprint Contract
Every task gets a BRIEF.md. The Sprint Contract section is mandatory — it lists specific, testable completion criteria.
# Task Brief
## Background
[Why this task exists]
## Objective
[What to build/fix]
## Sprint Contract (Completion Criteria)
- [ ] Criterion 1 (specific, verifiable)
- [ ] Criterion 2
- [ ] ...
⚠️ Write criteria specific to THIS task. No generic checklists.
## Related Files
[File paths relevant to the task]
## Constraints
[Tech stack, prior decisions, known pitfalls]
## Handoff Requirements
Write a HANDOFF.md when done, containing:
- What was done (file change list)
- Design decisions made (and why)
- What's left / known issues
- Everything needed for reporting to the human
2. Spawn Generator (Builder)
The generator receives the BRIEF.md and builds against the Sprint Contract. Key rules for the generator prompt:
- Work against the Sprint Contract criteria
- Self-check each criterion before handing off
- Write HANDOFF.md when done
- Write files first, read references second (output > research)
3. Spawn Evaluator (Independent QA)
After the generator finishes, spawn a separate agent as evaluator. The evaluator prompt must include:
The Sprint Contract — copied from BRIEF.md, to verify each criterion.
4 Evaluation Dimensions (select what's relevant):
| Dimension | What to check |
|---|---|
| Functional completeness | Every Sprint Contract criterion passes |
| User experience | Flow is intuitive, no dead ends |
| Visual quality | Layout, spacing, colors are professional |
| Code/content quality | No errors, clean logic, no regressions |
The critical prompt line:
"Your job is to find problems, not to praise. If everything looks fine, you probably didn't test carefully enough. Report issues honestly — better a false alarm than a missed bug."
4. Decision Gate
Based on evaluator feedback:
- All criteria pass → Ship it
- Criteria fail → Feed evaluator report back to generator for fixes
- Architecture issues → Escalate to human
When to Use Each Mode
| Task complexity | Generator | Evaluator | Example |
|---|---|---|---|
| Simple (< 30 min) | Sub-agent | Self-evaluate, mark "⚠️ untested" | Fix a typo, update config |
| Medium (30 min - 2 hr) | Sub-agent | Independent sub-agent | New feature, bug fix |
| Complex (2+ hr) | Claude Code / ACP | Independent sub-agent + human review | Architecture change, new project |
Sprint Contract Examples
See references/contract-examples.md for project-specific contract templates.
Key Insights from Anthropic's Research
- File-based communication — Agents talk through files (BRIEF.md, HANDOFF.md), not conversation
- Evaluator calibration — Default LLMs are too lenient; explicitly prompt for skepticism
- Sprint scoping — One feature at a time; don't bundle unrelated work
- Opus 4.6 + 1M context — Context anxiety is gone; sprint decomposition is less critical, but evaluator still adds value at task boundaries
- Evaluation criteria shape output — The wording of your criteria directly steers what the generator produces