Skip to content

Orchestration Patterns

Four patterns for combining multiple skills and agents into larger workflows.

Pattern 1: Sequential Pipeline

Run skills one after another, where each step's output feeds the next.

Code Review → Security Scan → Performance Profile → Release Readiness

Example: Pre-release quality gate

# Step 1: Code quality
python engineering/code-reviewer/scripts/code_quality_analyzer.py .

# Step 2: Security check
python engineering/senior-security/scripts/security_scanner.py .

# Step 3: Performance check
python engineering/performance-profiler/scripts/profiler.py .

# Step 4: Release readiness
python engineering/release-manager/scripts/release_readiness.py .

When to use: Linear workflows where order matters and each step validates a different dimension.

Pattern 2: Fan-Out / Fan-In

Run multiple skills in parallel, then merge results.

                ┌─ SEO Audit ──────┐
Input Page ─────┼─ A11y Audit ─────┼─── Merged Report
                └─ Performance ────┘

Example: Comprehensive page audit

Run three independent audits on the same page, then combine findings into a single report sorted by severity.

When to use: Independent analyses that can run concurrently. Reduces total wall-clock time.

Pattern 3: Agent Delegation

A top-level agent delegates subtasks to specialized agents.

cs-tech-lead
├── delegates to cs-code-auditor (code quality)
├── delegates to cs-security-engineer (security)
└── synthesizes findings into architecture review

Example: The chief-of-staff skill routes founder questions to the right C-suite advisor (CEO, CTO, CFO, etc.) and synthesizes their responses.

When to use: Complex tasks that span multiple domains and need a coordinator.

Pattern 4: Iterative Refinement

Run a skill, evaluate the output, and re-run with adjustments.

Draft Content → SEO Score → Revise → SEO Score → Publish

Example: Content production loop

  1. Generate draft with content-creator
  2. Score with seo-specialist tools
  3. If score < threshold, revise and re-score
  4. Publish when quality gate passes

When to use: Tasks where quality improves through iteration and you have a measurable quality metric.

Choosing a Pattern

Pattern Best For Complexity
Sequential Quality gates, release pipelines Low
Fan-Out/Fan-In Multi-dimensional analysis Medium
Agent Delegation Cross-domain strategy Medium
Iterative Refinement Content, optimization High

Start simple

Most workflows only need Pattern 1 (sequential). Add complexity only when you have a clear reason.

Building Custom Workflows

Combine patterns using shell scripts, GitHub Actions, or the agenthub skill for DAG-based orchestration:

# See the agenthub skill for multi-agent DAG orchestration
cat engineering/agenthub/SKILL.md