Multi-Agent Workflow

Orchestrate complex tasks across multiple AI agents, each contributing their strengths.

Agent Roles

Workflow Patterns

Pattern 1: Sequential Pipeline

Each agent builds on the previous one's output:

Gemini (Research) → Claude (Implement) → Codex (Review) → Claude (Fix)

Use when: Tasks have clear dependencies between phases

Example workflow:

1. Gemini researches the API/library to use
2. Claude implements based on research findings
3. Codex reviews for bugs and security
4. Claude fixes any issues found

Pattern 2: Parallel Specialization

Multiple agents work on different aspects simultaneously:

         ┌─ Claude (Bug Review)
Task ────┼─ Gemini (Security Review)
         └─ Codex (Performance Review)

Use when: Independent reviews or analyses can happen concurrently

Example workflow:

1. Send the same code to all three agents
2. Each focuses on their specialty
3. Combine findings at the end

Pattern 3: Supervisor-Worker

One agent coordinates others:

Claude (Supervisor)
    ├─ delegates to → Gemini (Research subtask)
    ├─ delegates to → Codex (Test subtask)
    └─ integrates results

Use when: Complex tasks need coordination and integration

Implementation Steps

Step 1: Plan the Workflow

Define the task and identify which agents should participate:

const workflow = {
  task: "Implement user authentication",
  phases: [
    { role: "research", agent: "gemini-cli", input: "Research OAuth2 best practices" },
    { role: "implement", agent: "claude-code", input: "Implement based on research" },
    { role: "test", agent: "codex-cli", input: "Write and run security tests" },
    { role: "fix", agent: "claude-code", input: "Fix any issues found" }
  ]
};

Step 2: Execute with Handoffs

Use cliagents' delegation tools:

delegate_task role="research" adapter="gemini-cli" message="Research OAuth2..."
delegate_task role="implement" adapter="claude-code" message="[Research results] Implement..."
delegate_task role="test" adapter="codex-cli" message="[Implementation] Test..."

Step 3: Share Context

Use shared memory to pass information between agents:

• store_artifact: Save implementation plans, code outputs
• share_finding: Report bugs, security issues, suggestions
• get_shared_findings: Retrieve what other agents discovered

Step 4: Integrate Results

Combine outputs from all agents into the final deliverable.

Best Practices

1. Clear handoff messages: Include all context the next agent needs
2. Use shared memory: Store artifacts and findings for other agents
3. Parallel when possible: Use parallel patterns for independent work
4. Single source of truth: Let one agent own the final integration
5. Validate incrementally: Check each agent's output before proceeding

Predefined Workflows

Use run_workflow for common patterns:

• code-review: Parallel bug + security + performance review
• feature: Plan → implement → test
• bugfix: Analyze → fix → test
• research: Research → document

Error Handling

• If an agent times out: Increase timeout or use async mode
• If an agent fails: Retry or fall back to another agent
• If agents disagree: Escalate to supervisor agent for resolution