workflow-patterns_skill

Overview

This skill provides a technology-agnostic catalog of workflow patterns and coordination strategies for orchestrating multiple agents. It summarizes core patterns, workflow templates, coordination primitives, error handling and anti-patterns to help designers pick the right orchestration model. Use it to standardize how agents hand off work, track progress, and recover from failures.

How this skill works

The skill describes common patterns (sequential chain, parallel execution, fork-join, iteration, conditional routing) and gives concrete implementation options such as wait-for-completion, aggregation, and fail-fast. It provides workflow templates (standard implementation, full feature, debug loop), progress-tracking schemas, handoff protocols, and error-handling rules to enforce coordination. Use the patterns to map tasks to agents, define exit conditions, and prevent deadlocks or infinite loops.

When to use it

• Designing an end-to-end multi-agent feature delivery pipeline
• Coordinating parallel tasks that must later be combined (fork-join)
• Implementing retry or refinement cycles with clear exit conditions (iteration)
• Routing work based on evaluation results or agent verdicts (conditional routing)
• Enforcing progress tracking, handoffs, and standardized error handling

Best practices

• Match orchestration complexity to task scale to avoid over-orchestration
• Always define exit conditions and max iterations to prevent infinite loops
• Use clear handoff summaries so the receiving agent has context and prerequisites
• Track per-task and overall progress with timestamps and status fields
• Prefer fail-fast for independent parallel tasks when time is critical; aggregate when results must be combined

Example use cases

• Sequential feature development: clarifier → architect → developer → tester with stop-on-failure
• Parallel module development: launch multiple code-developers and aggregate outputs
• Fork-join reviews: run security, performance, and style checks then combine results for evaluation
• Debug loop: repeat test → analyze → fix until tests pass or max iterations reached
• Conditional routing: route deliverables to commit or rework based on evaluator verdict