Documentation/Bushido/Bushido/ensure-subagent
📄

ensure-subagent

Hook File

Ensure Proper Subagent Usage

When to Use Subagents

Use Task tool for:

  • ✅ Parallel operations (exploration, review, analysis)
  • ✅ Extensive codebase search/exploration
  • ✅ Complex work needing focused attention
  • ✅ Multiple independent perspectives

❌ Don't use for simple operations - use Read/Grep/Glob instead.

Core Principles

1. Complete Task Descriptions

Provide autonomous instructions - subagents can't ask questions.

Include:

  • Clear objective and scope
  • Expected output format
  • Confidence thresholds (≥80%)
  • Specific exclusions (pre-existing issues, linter-caught problems)

2. Parallel Execution (CRITICAL)

Always launch independent agents in a SINGLE message:

Single message with multiple Task calls:
- Agent 1: [task A]
- Agent 2: [task B]
- Agent 3: [task C]

❌ Never sequential when tasks are independent.

3. Confidence-Based Filtering

All review findings must include confidence score (0-100):

  • ≥90%: Critical issues, report always
  • ≥80%: Important issues, report
  • <80%: Filter out, don't report

Exclude from reports:

  • Pre-existing issues (not in current changes)
  • Linter-catchable problems
  • Style preferences without standards
  • Theoretical concerns without evidence

4. Consolidation

When running multiple agents:

  1. Collect all findings
  2. De-duplicate identical issues
  3. Filter for confidence ≥80%
  4. Resolve conflicts (highest confidence wins, security overrides)
  5. Present unified report

5. Human Decision Points

Use AskUserQuestion for:

  • Multiple valid approaches with trade-offs
  • Unclear requirements
  • Breaking changes affecting others
  • Long-term technical decisions

Don't pause for:

  • Standard documented patterns
  • Obviously correct approaches
  • Minor implementation details

Quick Rules

Do:

  • Launch parallel agents in single message
  • Trust subagent expertise
  • Provide complete context in prompts
  • Use appropriate subagent_type (Explore/Plan/General/Specialized)
  • Filter all findings to ≥80% confidence

Don't:

  • Use subagents for simple operations
  • Launch agents sequentially when parallel possible
  • Spawn agents for work you're already doing
  • Report low-confidence findings
  • Second-guess domain experts without reason
Back to Bushido