agentic-workflow-orchestration（Agent 工作流编排）

🎭 Agentic Workflow Orchestration Skill

Purpose

Guidance on orchestrating multiple agentic workflows for complex problems through coordination, delegation, and parallel execution. Includes proven patterns from the Agent Factory (100+ workflows, 500+ merged PRs) including /plan decomposition, task mining, and stacked agent coordination.

When to Use

Apply this skill when:

• Breaking down complex automation into coordinated workflows
• Implementing orchestrator-worker patterns
• Using /plan command for issue decomposition
• Building pipeline or map-reduce patterns
• Coordinating multiple AI agents with different specializations

Rules

Orchestration Fundamentals

MUST:

• Design orchestrators to coordinate, not execute
• Create focused workers with single responsibilities
• Use workflow dispatch for orchestrator-worker communication
• Implement clear contracts between orchestrator and workers
• Handle worker failures gracefully in orchestrators
• Aggregate and validate worker results
• Log orchestration decisions and outcomes

MUST NOT:

• Mix orchestration logic with execution logic
• Create overly complex orchestration hierarchies
• Ignore worker failures
• Create circular dependencies between workflows
• Skip result validation from workers

Orchestrator Design

MUST:

• Keep orchestrators stateless when possible
• Use GitHub APIs to track worker status
• Implement retry logic for failed workers
• Provide clear success/failure criteria
• Generate summary reports of orchestrated work
• Scale workers horizontally, not vertically
• Document orchestration logic clearly

MUST NOT:

• Hard-code worker names or configurations
• Assume workers will always succeed
• Create long-running orchestrators (use scheduled checks instead)
• Tightly couple orchestrators to specific worker implementations

Worker Design

MUST:

• Focus each worker on single, well-defined task
• Make workers independently testable
• Report clear success/failure status
• Provide detailed output for orchestrator consumption
• Handle edge cases locally
• Implement appropriate timeouts
• Use consistent output formats

MUST NOT:

• Create workers with multiple responsibilities
• Depend on other workers directly (use orchestrator)
• Assume orchestrator context
• Create workers that never finish
• Skip error reporting

Communication Patterns

MUST:

• Use workflow_dispatch with inputs for orchestrator→worker communication
• Use issues, comments, or pull requests for worker→orchestrator communication
• Implement structured data formats (JSON, YAML) for inter-workflow communication
• Version communication protocols
• Validate all inter-workflow messages
• Document communication contracts

MUST NOT:

• Use file system for inter-workflow communication
• Trust unchecked data from other workflows
• Create hidden communication channels
• Skip message validation

Parallel Execution

MUST:

• Dispatch independent workers in parallel
• Set appropriate concurrency limits
• Implement timeout handling for parallel workers
• Collect and validate results from all workers
• Handle partial success scenarios
• Track execution metrics (timing, success rates)

MUST NOT:

• Wait for workers sequentially when parallel execution is possible
• Create unbounded parallelism
• Ignore partial failures
• Assume workers complete in specific order

Error Handling and Resilience

MUST:

• Implement circuit breakers for failing workers
• Use exponential backoff for retries
• Set maximum retry limits
• Log all failures with context
• Escalate persistent failures to humans
• Implement graceful degradation
• Maintain audit trail of orchestration decisions

MUST NOT:

• Retry indefinitely
• Hide failures from orchestrators
• Cascade failures without containment
• Skip logging of error conditions

Orchestration Patterns

Pattern 1: Simple Orchestrator-Worker

Structure:

• 1 orchestrator workflow
• N homogeneous worker workflows
• Workers execute identical tasks on different inputs

Use Cases:

• Batch processing (analyze all open issues)
• Parallel validation (test multiple configurations)
• Distributed scanning (security audit across repos)

Pattern 2: Hierarchical Orchestration

Structure:

• Top-level orchestrator
• Mid-level specialized orchestrators
• Leaf-level execution workers

Use Cases:

• Complex multi-phase operations
• Different orchestration strategies per phase
• Nested task decomposition

Pattern 3: Pipeline Orchestration

Structure:

• Orchestrator manages sequential stages
• Each stage has specialized workers
• Output of one stage feeds next stage

Use Cases:

• CI/CD pipelines
• Data processing pipelines
• Multi-step validation workflows

Pattern 4: Router/Dispatcher

Structure:

• Router analyzes input and selects appropriate worker
• Heterogeneous workers with different capabilities
• Dynamic worker selection based on context

Use Cases:

• Issue triage (route to specialized handlers)
• Content moderation (route by content type)
• Multi-language support (route by language)

Pattern 5: Map-Reduce

Structure:

• Map phase: Orchestrator spawns workers for each input
• Reduce phase: Orchestrator aggregates worker results
• Final report generation

Use Cases:

• Repository analysis across multiple repos
• Parallel testing with result aggregation
• Distributed data collection and synthesis

Pattern 6: Reflection (Self-Improving)

Structure:

• Generator workflow produces output
• Critic workflow reviews and provides feedback
• Coordinator orchestrates iterations until quality threshold met

Use Cases:

• Documentation generation with quality review
• Code generation with security review
• Content creation with style validation

Examples

Example 1: Simple Orchestrator-Worker Pattern

Orchestrator (daily-orchestrator.md):

---
on: daily
permissions: read-all
tools:
  github:
safe-outputs:
  create-issue:
    max: 1
---

# Daily Repository Orchestrator

Coordinate daily analysis across all open issues:

1. List all open issues in repository
2. For each issue, dispatch `issue-analyzer` worker workflow with issue number
3. Wait for all workers to complete (check workflow run statuses)
4. Collect results from worker comments
5. Generate summary report issue titled "Daily Analysis Report - [DATE]"

Include:
- Total issues analyzed
- Critical issues found
- Recommendations summary
- Links to worker analysis comments

Worker (issue-analyzer.md):

---
on:
  workflow_dispatch:
    inputs:
      issue_number:
        required: true
        type: number
permissions: read-all
tools:
  github:
safe-outputs:
  create-comment:
    max: 1
---

# Issue Analyzer Worker

Analyze issue #${{ inputs.issue_number }}:

1. Read issue title, body, labels, and comments
2. Check for duplicates
3. Assess clarity and completeness
4. Identify missing information
5. Suggest appropriate labels

Post analysis as comment on the issue.
Use structured format:
- Status: [Clear/Unclear/Duplicate]
- Suggested Labels: [label1, label2]
- Action: [what should happen next]

Example 2: Hierarchical Orchestration

Top-Level Orchestrator (repo-health-orchestrator.md):

---
on: weekly on monday
permissions: read-all
tools:
  github:
safe-outputs:
  create-issue:
    max: 1
---

# Repository Health Orchestrator

Coordinate comprehensive repository health check:

Phase 1: Dispatch specialized orchestrators
1. Dispatch `issue-health-orchestrator`
2. Dispatch `pr-health-orchestrator`
3. Dispatch `code-health-orchestrator`

Phase 2: Wait and collect
1. Wait for all phase 1 orchestrators to complete
2. Collect results from each orchestrator issue

Phase 3: Synthesize
1. Combine results into master health report
2. Generate action items by priority
3. Create issue with complete repository health status

Mid-Level Orchestrator (issue-health-orchestrator.md):

---
on:
  workflow_dispatch:
permissions: read-all
tools:
  github:
safe-outputs:
  create-issue:
    max: 1
---

# Issue Health Orchestrator

Coordinate issue-specific analysis:

1. Dispatch workers for:
   - Stale issue detection
   - Issue categorization
   - Response time analysis
   - Label consistency check

2. Collect worker results
3. Generate issue health report
4. Create issue with findings

Example 3: Pipeline Orchestration

Pipeline Orchestrator (deployment-pipeline.md):

---
on: workflow_dispatch
permissions: read-all
tools:
  github:
safe-outputs:
  create-comment:
    max: 5
---

# Deployment Pipeline Orchestrator

Coordinate deployment pipeline stages:

Stage 1: Validation
1. Dispatch `security-validator` worker
2. Dispatch `test-validator` worker
3. Wait for both, check results
4. If any failures, stop and report

Stage 2: Staging Deployment
1. Dispatch `staging-deployer` worker
2. Wait for completion
3. Dispatch `staging-validator` worker
4. If validation fails, rollback and stop

Stage 3: Production Deployment
1. Require manual approval (create approval issue)
2. After approval, dispatch `production-deployer` worker
3. Dispatch `production-validator` worker
4. Generate deployment report

Post progress comments at each stage.

Example 4: Router/Dispatcher Pattern

Router (issue-router.md):

---
on: issues
permissions: read-all
tools:
  github:
---

# Issue Router

Route new issues to specialized handlers:

1. Analyze issue labels and content
2. Determine issue type:
   - Bug → dispatch `bug-triage-worker`
   - Feature Request → dispatch `feature-triage-worker`
   - Security → dispatch `security-triage-worker`
   - Documentation → dispatch `docs-triage-worker`
   - General → dispatch `general-triage-worker`

3. Dispatch appropriate worker with issue number
4. Post comment indicating routing decision

Example 5: Map-Reduce Pattern

Map-Reduce Orchestrator (security-audit-orchestrator.md):

---
on: workflow_dispatch
permissions: read-all
tools:
  github:
safe-outputs:
  create-issue:
    max: 1
  create-code-scanning-alert:
    max: 1
---

# Security Audit Orchestrator (Map-Reduce)

Coordinate security audit across entire codebase:

MAP PHASE:
1. List all repositories in organization
2. For each repository, dispatch `security-scanner` worker
3. Track dispatched workers

REDUCE PHASE:
1. Wait for all scanners to complete
2. Collect SARIF reports from each worker
3. Aggregate findings:
   - Total vulnerabilities by severity
   - Top vulnerability types
   - Repositories with most issues
4. Generate consolidated SARIF report
5. Create master security issue with summary

Example 6: Reflection Pattern

Reflection Orchestrator (doc-generator-orchestrator.md):

---
on: workflow_dispatch
permissions: read-all
tools:
  github:
safe-outputs:
  create-pull-request:
    max: 1
---

# Documentation Generator with Reflection

Use reflection pattern for high-quality docs:

Iteration Loop (max 3 iterations):
1. Dispatch `doc-generator` worker to create documentation
2. Dispatch `doc-critic` worker to review generated docs
3. Collect critique feedback
4. If quality score >= 8/10, proceed to next step
5. If quality score < 8/10, regenerate with critique feedback

Final Step:
1. Create pull request with approved documentation
2. Include quality metrics in PR description
3. Link to critic reviews

Example 7: Circuit Breaker Pattern

Orchestrator with Circuit Breaker (resilient-orchestrator.md):

---
on: schedule
permissions: read-all
tools:
  github:
safe-outputs:
  create-issue:
    max: 1
---

# Resilient Orchestrator with Circuit Breaker

Coordinate workers with failure protection:

For each task in task_list:
  1. Check circuit breaker state for worker type
  2. If OPEN (too many recent failures), skip and log
  3. If CLOSED, dispatch worker
  4. Track worker success/failure
  5. Update circuit breaker state:
     - 3 consecutive failures → OPEN circuit
     - 1 success after OPEN → HALF-OPEN
     - 3 successes in HALF-OPEN → CLOSED

Generate report with:
- Tasks completed successfully
- Tasks skipped (circuit open)
- Circuit breaker states
- Recommended actions

Example 8: Dynamic Worker Pool

Dynamic Pool Orchestrator (scalable-orchestrator.md):

---
on: workflow_dispatch
permissions: read-all
tools:
  github:
safe-outputs:
  create-issue:
    max: 1
---

# Dynamic Worker Pool Orchestrator

Scale workers based on workload:

1. Assess workload size (e.g., count of items to process)
2. Calculate optimal worker count (workload / items_per_worker)
3. Respect concurrency limits (max 10 concurrent workers)
4. Batch items for each worker
5. Dispatch workers with batch assignments
6. Monitor worker progress
7. If workers fail, redistribute work to new workers
8. Aggregate results when all batches complete
9. Generate completion report with metrics:
   - Items processed
   - Workers used
   - Success rate
   - Processing time

Best Practices

Orchestrator Design

• Keep orchestration logic simple and linear
• Use timeouts to prevent infinite waiting
• Implement health checks for long-running coordination
• Log all orchestration decisions
• Generate human-readable summary reports

Worker Design

• Make workers idempotent (safe to retry)
• Include worker ID in outputs for traceability
• Report progress for long-running workers
• Use consistent success/failure signals
• Minimize external dependencies

Communication

• Use structured data (JSON/YAML) for all inter-workflow data
• Version communication protocols
• Validate all inputs from other workflows
• Document communication contracts clearly

Testing

• Test workers independently before orchestration
• Test orchestrator with mock worker responses
• Test failure scenarios (timeouts, errors)
• Test partial success scenarios
• Load test with realistic workloads

Monitoring

• Track orchestration execution time
• Monitor worker success rates
• Alert on orchestration failures
• Measure end-to-end latency
• Dashboard key metrics

Troubleshooting

Workers Not Completing

• Check worker timeout settings
• Review worker logs for errors
• Verify worker dispatch succeeded
• Check GitHub Actions queue status
• Implement worker health checks

Orchestration Hangs

• Verify timeout logic in orchestrator
• Check for circular dependencies
• Review worker status polling logic
• Ensure failure cases are handled

Inconsistent Results

• Verify worker idempotency
• Check for race conditions
• Review result aggregation logic
• Validate worker communication format

Performance Issues

• Optimize worker batch sizes
• Increase parallelism if possible
• Review orchestrator polling frequency
• Consider worker caching strategies

Related ISMS Policies

This skill aligns with:

• Secure Development Policy - Secure orchestration practices
• Change Management Policy - Coordinated deployment patterns
• Information Security Policy - Overall security framework

Related Skills

• GitHub Agentic Workflows - Core workflow fundamentals
• Agentic Workflow Security - Security for orchestrated workflows
• Agentic Workflow Development - Development and testing
• Copilot Agent Patterns - Custom agent patterns

Related Documentation

• Agent Factory Workflows
• Orchestration Guide
• GitHub Actions Reusable Workflows

Compliance Mapping

ISO 27001:2022

• A.8.25 Secure development life cycle
• A.8.32 Change management
• A.5.37 Documented operating procedures

NIST Cybersecurity Framework 2.0

• PR.IP-02: A System Development Life Cycle is implemented
• PR.MA-01: Maintenance is performed
• DE.AE-03: Event data is collected and correlated

CIS Controls v8.1

• Control 16: Application Software Security

- 16.1 Establish Secure Development Process - 16.14 Establish Remediation Processes

Enforcement

Orchestration pattern violations:

• Critical: Circular dependencies, infinite loops - Block deployment
• High: Missing error handling, no timeouts - Require remediation
• Medium: Suboptimal patterns, missing monitoring - Create improvement tickets
• Low: Documentation gaps, optimization opportunities - Optional improvements

Version History

• 2026-04-02: Updated with Agent Factory proven patterns (/plan, task mining), real merge rate data
• 2026-02-11: Initial skill creation