Competitive Agents

Overview

Dispatch two independent subagents to solve the same task from different angles. The main agent then synthesizes the best elements into a superior combined result.

When to Use

• Designing systems, architectures, or APIs
• Problems with multiple valid approaches (no single obvious answer)
• User explicitly requests competitive/parallel approaches
• You want to reduce single-agent bias

When NOT to Use

• Purely mechanical tasks (rename variable, fix typo, add import)
• Tasks requiring sequential steps (not parallelizable)
• Task is unclear — clarify with user first, then decide

Red Flags - Skill Still Applies

Don't skip just because:

• Task seems "simple" (simplicity ≠ obvious solution)
• You think one approach is "best" (that's exactly the bias to avoid)

If multiple valid approaches exist, use this skill.

Workflow

1. Clarify if needed — if the task is ambiguous, ask the user before dispatching
2. Dispatch 2 subagents in parallel (single message, 2 Task tool calls)
3. Wait for both results
4. You (main agent) synthesize the best elements directly — no third subagent
5. Present synthesized result to user

Competitor Prompt Template

Each competitor gets the same task but a different constraint to force divergent solutions:

Competitor A:

You are competing against another agent to solve the same task.
The better solution will be selected. Give your absolute best effort.

YOUR CONSTRAINT: Prioritize simplicity and minimalism. Fewer moving parts wins.

## Task
{task content from user}

## Requirements
- Provide a complete, well-reasoned solution
- Explain your approach and key decisions
- Consider trade-offs and alternatives you rejected

Competitor B:

You are competing against another agent to solve the same task.
The better solution will be selected. Give your absolute best effort.

YOUR CONSTRAINT: Prioritize completeness and extensibility. Cover more cases and future needs.

## Task
{task content from user}

## Requirements
- Provide a complete, well-reasoned solution
- Explain your approach and key decisions
- Consider trade-offs and alternatives you rejected

Both use subagent_type: "general-purpose".

Synthesis (Main Agent)

After both competitors complete, you synthesize directly. No subagent needed — you already have both results in context.

Analyze and present in this format:

Analysis

Solution A: [strengths] / [weaknesses] Solution B: [strengths] / [weaknesses]

Synthesized Solution

[combined best solution]

Rationale

[what you took from each and why]

Execution

Dispatch both competitors in a single message (parallel):

Task call 1: description="Competitor A - simplicity", subagent_type="general-purpose"
Task call 2: description="Competitor B - completeness", subagent_type="general-purpose"

After both return, synthesize directly as the main agent (see Synthesis section above).

Common Mistakes

Mistake	Fix
Both competitors produce identical solutions	Use the constraint prompts above to force divergence
Synthesizing before both competitors complete	Use foreground mode, wait for both
Skipping synthesis	Always synthesize — even if one seems clearly better
Dispatching a judge subagent	Main agent synthesizes directly — no third Task call needed
Dispatching on an unclear task	Clarify with user first (workflow step 1)
Skipping because task seems "simple"	Simplicity of description ≠ obvious solution