parallel-analyze

Parallel Analyze

Split the same task across parallel agents (one per target), then merge their results into a single report.

When to Use This Pattern

• The same task applies to multiple independent targets (audit 5 modules, review 8 files, analyze 12 endpoints)
• Each target can be processed independently
• You need a combined summary, not just individual results

Process

Phase 1: Map (Identify Targets)

List all targets. For each target, confirm:

1. It can be processed independently (no cross-dependencies)
2. The same prompt and criteria apply to all targets
3. The expected output format is consistent

Phase 2: Map (Dispatch)

Spawn one agent per target (or per batch of small targets). All agents:

• Receive the same instructions and criteria
• Get their specific target(s) as context
• Use the same restricted tool set
• Return results in the same format

Launch all agents in parallel (multiple Agent calls in one message).

Phase 3: Reduce (Merge)

When all agents return:

1. Collect all results
2. Deduplicate findings that appear across multiple targets
3. Rank by severity or importance
4. Produce a single summary with:
   • Overall assessment
   • Per-target highlights (only notable items, not everything)
   • Cross-cutting patterns (issues that appear in 3+ targets)
   • Recommended actions, prioritized

Example

Task: "Security audit all API route handlers"

Map: 6 route files, each gets its own agent with the same security checklist. Reduce: Merge into one report. Flag patterns ("4 of 6 routes missing rate limiting") over individual findings.

Sizing Guide

Targets	Strategy
1-2	Do not use parallel-analyze. Just do the work.
3-6	One agent per target. Ideal range.
7-15	Batch targets into 3-5 groups. One agent per group.
15+	Batch into 5-7 groups. Consider if the task scope is too broad.

Rules

• All map agents must use the same output format. Define it in the prompt.
• Cap at 7 parallel agents. More than that and the reduce phase becomes unwieldy.
• The reduce step is your job, not another agent's. You have the full picture.
• If one agent fails or returns poor results, re-run just that one.