Subagent Workflow Patterns To Boost Output Quality

Orchestrate Multi-Agent Workflows Like a Team Lead

Stop struggling with "agentic laziness" and goal drift in long conversations. This skill transforms your AI agent into a sophisticated orchestrator capable of managing teams of specialized subagents through six battle-tested architectural patterns.

The 6 Patterns — With Real Examples

🔀 Fan-Out & Synthesize

Parallelize massive tasks across multiple subagents. Instead of researching one competitor at a time, spawn 3 agents simultaneously — each analyzes a different competitor — then merge their findings into one competitive landscape report. Perfect for code reviews (20+ files in parallel), market research, and migrations.

🛡️ Adversarial Verification

Solve self-bias by spawning a "skeptic" agent to find flaws before delivery. Your agent writes code or content → a harsh reviewer tears it apart against a rubric → only passes when the critic is satisfied. No more "looks good to me" when it isn't. Perfect for security audits, fact-checking, and quality-critical deliverables.

🏆 Tournament

Generate multiple competing solutions and use a pairwise judge to pick the objective winner. Need to choose a pricing model? 3 agents each argue for a different approach → judge compares A vs B → winner faces C → champion emerges. Better than picking the first idea that comes to mind.

🧠 Generate-and-Filter

Produce dozens of ideas across parallel streams, then distill them into the top 1%. Brainstorming a store name? 3 agents generate 10 names each from different angles → a filter agent scores, ranks, and deduplicates → you get the top 5 winners. Quantity into quality.

🔀 Classify-and-Act

Intelligently route complex requests to the right specialist subagent. Incoming bug report? A classifier quickly determines if it's frontend, backend, or infrastructure — then delegates to the appropriate fixer. No more wasting time with the wrong specialist.

🔄 Loop-Until-Done

Ensure 100% completion for exhaustive tasks. Need to fix all SEO tags across your store? Spawn an agent to find and fix missing meta descriptions → scan again → fix more → repeat until zero findings remain. No more "I fixed most of them."

What's Included in This Skill

• SKILL.md — Full configuration file with all 6 patterns and domain mappings
• 6 Template Files — One ready-to-use prompt template per pattern (fan-out, adversarial, tournament, classify, generate-filter, loop)
• Domain Quick-Reference — Which patterns to reach for in software development, content creation, business strategy, learning & research, creative design, and ecommerce
• Sample Inputs & Outputs — Real worked examples showing exactly what each pattern produces

Why Use This Skill

Individual agents often lose focus or settle for "good enough." By using these orchestration patterns, you enforce a rigorous process that mimics high-performing human teams — parallel workers, quality reviewers, judges, and loopers. No external dependencies or complex infrastructure required.

Before vs After

• 😞 Before: Agent fixes 20/50 bugs and stops → ✅ After: Loop-Until-Done ensures all 50 are fixed and verified
• 😞 Before: Agent says "looks good" to its own code → ✅ After: Adversarial agent finds real flaws every time
• 😞 Before: You manually split work across chat tabs → ✅ After: Fan-out spawns parallel agents — one merged result
• 😞 Before: Agent picks the first solution it thought of → ✅ After: Tournament pits 3 approaches, best one wins

Supported Tools

Optimized for Claude Code, Cursor, Aider, Windsurf, Codex CLI, Gemini CLI, GitHub Copilot, and any agent supporting the SKILL.md standard or MCP-based subagent delegation.