dispatching-parallel-agents

Dispatching Parallel Agents

Register: TECHNIQUE Goal: Solve multiple independent problems in parallel time without context contamination between agents. Constraints: True independence required — no shared files, no sequential dependencies. Each agent gets its own clean context. pattern-globalize after any agent finds a bug. Adapt: number of agents, context depth, and coordination protocol to task complexity and available token budget.

---

When to Use

Use when ALL are true:

• 2+ independent failures or tasks (different subsystems, different test files)
• Each can be understood without context from the others
• Agents won't touch the same files

Don't use when:

• Failures might be related (fixing one may fix others — investigate first)
• Need full system context to understand the problem
• Agents would touch shared files

---

The Pattern

1. Identify Independent Domains

Goal: group failures by structural independence, not by file or feature proximity.

Domain A: auth module — 3 test failures
Domain B: data pipeline — 2 test failures
Domain C: API handlers — 1 test failure

Each domain independent: auth failure won't affect pipeline behavior.

2. Construct Focused Agent Context

Goal: give each agent exactly what it needs — no more, no less. Constraints:

• Never let subagents inherit your session context — they get confused
• Specify scope explicitly: "Do NOT change files outside tests/auth/"
• Specify output format: "Return: root cause summary + list of file:line changes"

Fix the N failing tests in <path>:

1. "<test name>" — <expected behavior from test>
2. "<test name>" — <expected behavior from test>

Your task:
1. Read the test file — understand what each test verifies
2. Find root cause using systematic-debugging Phase 1-3
3. Fix using tdd-verified (failing test → implement → passing test)
4. Do NOT change files outside <your scope>

Return: root cause summary + changes made (file:line for each)