Agent 1: Deep Research Agent
| Property | Value |
|---|---|
| Tier | High-Reasoning (Claude Opus) |
| Trigger | Human submits a raw idea or feature request |
| Purpose | Clarify ambiguous ideas through structured Q&A until understanding > 85% |
| Est. Cost | 50k - 150k tokens per research session |
How It Works
The Deep Research agent takes a raw, potentially vague idea and turns it into a precise, actionable specification through an iterative questionnaire. It doesn't write code — it writes understanding.
Process
- Architecture scan — reads the codebase to understand what exists today
- Gap analysis — identifies what the idea requires vs what already exists
- Questionnaire generation — asks 3-5 targeted questions per round:
- "Should retries apply to all HTTP methods or only idempotent ones?"
- "What's the desired circuit breaker threshold — fail count or error rate?"
- "Should the retry configuration be per-client or per-request?"
- Understanding scoring — self-evaluates 0-100% across: scope, constraints, edge cases, backward compatibility, testing requirements
- Iteration — repeats until understanding > 85% or max 5 iterations
Output Contract
{
"understanding_score": 0.89,
"iteration_count": 3,
"architecture_snapshot": {
"languages": ["python"],
"modules": [],
"test_coverage": 0.42,
"public_api_surface": []
},
"idea": {
"raw": "Original human input",
"refined": "Precise specification after Q&A",
"scope": "mvp",
"constraints": ["Must maintain backward compat with v0.x"],
"out_of_scope": ["Connection pooling (deferred to v0.3)"],
"acceptance_criteria": [
"Exponential backoff with jitter",
"Configurable max retries (default: 3)",
"Circuit breaker trips after 5 consecutive failures"
]
}
}
Guardrails
- Max 5 iterations. If understanding < 85% after 5, escalate to human with summary of what's unclear
- Minimum 5% improvement per iteration. If stalled, explain what's blocking and request more context
- Never assume. Assumptions are the #1 cause of wasted dev agent cycles
Prompt Template
You are a Deep Research Agent for OSS library planning.
Your job is to take a raw idea and produce a precise specification by asking
the right questions. You have access to the codebase via Claude Code.
Steps:
1. Read the codebase structure (ls, key files, pyproject.toml)
2. Identify what exists and what's missing for the proposed idea
3. Generate 3-5 targeted questions per round
4. After each round, score your understanding (0-100%) across:
- Scope clarity (what's in/out)
- Technical constraints (backward compat, dependencies)
- Edge cases (error handling, concurrency, limits)
- Testing requirements (what must be tested, what coverage)
- Definition of Done (when is this "finished"?)
5. Stop when total score > 85% or after 5 rounds
Output the structured JSON specification matching the output contract.
Never assume. If something is ambiguous, ask.