---
name: implementation
description: Adaptive implementation workflow where a planner agent decides the subagent fanout
topology: complex-dag
---

## assess
role: planner
output: adaptive-plan.json

Assess this task and decide how many subagents are actually needed for: {goal}

You are the orchestration planner. Inspect the repository enough to choose an efficient crew; do not use a fixed template. Small/simple tasks may need one executor plus one verifier. Risky or broad tasks may need parallel explorers, specialists, implementers, reviewers, security reviewers, or test engineers.

Return a concise rationale, then include exactly one JSON block between these markers:

ADAPTIVE_PLAN_JSON_START
{
  "phases": [
    {
      "name": "short-phase-name",
      "tasks": [
        {
          "role": "explorer|analyst|planner|critic|executor|reviewer|security-reviewer|test-engineer|verifier|writer",
          "title": "short task title",
          "task": "specific autonomous task prompt for this subagent"
        }
      ]
    }
  ]
}
ADAPTIVE_PLAN_JSON_END

Rules:
- **MAXIMIZE PARALLELISM**: Put independent tasks in the SAME phase so they run concurrently.
  For example, if a task needs exploration + implementation + review, use 3 phases:
  Phase 1: explorers (2-3 in parallel), Phase 2: executors (2-3 in parallel), Phase 3: reviewers (2 in parallel).
  NEVER create sequential phases when tasks are independent.
- Choose the smallest effective number of subagents per phase.
- Tasks within the same phase run in parallel; phases run sequentially.
- Include verification/review tasks when implementation is requested.
- Do not include more than 12 total subagents; split or summarize oversized plans instead.
- A good plan for a complex task has 2-4 phases with 2-4 parallel tasks each.
- A simple task may have just 1-2 phases with 1-2 tasks.