​

Orchestration Teams

​

How It Works

User Input
    │
    ▼
┌──────────────┐        delegate_to_worker()
│  Brain Agent │ ──────────────────────────────► Worker A
│  (Orchestrator)│ ◄──── result ─────────────────
│              │
│              │        delegate_to_worker()
│              │ ──────────────────────────────► Worker B
│              │ ◄──── result ─────────────────
│              │
│              │  [quality < threshold?]
│              │        request_revision()
│              │ ──────────────────────────────► Worker A
│              │ ◄──── revised result ──────────
│              │
│              │        final_answer()
└──────────────┘
    │
    ▼
Final Output

​

Creating an Orchestration Team

1. Navigate to Teams → New Team
2. Give it a name and description
3. Select Orchestrator as the team mode
4. Configure the brain and worker agents (see below)

​

Configuration

​

Brain Agent

Tip: The brain’s system prompt is overridden by CrewForm’s orchestrator prompt, which teaches it how to use the delegation tools. The agent’s own system prompt is prepended as additional context.

​

Worker Agents

You are a senior TypeScript developer specialising in React and performance optimisation.
Review the code for bugs, type errors, and performance issues.

​

Configuration Fields

​

Brain Agent Tools

​

delegate_to_worker

{
  "tool": "delegate_to_worker",
  "arguments": {
    "agent_id": "<worker-agent-id>",
    "instruction": "Analyse the performance bottlenecks in the provided React component and suggest optimisations."
  }
}

​

request_revision

{
  "tool": "request_revision",
  "arguments": {
    "delegation_id": "<delegation-id>",
    "feedback": "The analysis is too generic. Focus specifically on unnecessary re-renders and memo opportunities."
  }
}

​

accept_result

{
  "tool": "accept_result",
  "arguments": {
    "delegation_id": "<delegation-id>"
  }
}

​

final_answer

{
  "tool": "final_answer",
  "arguments": {
    "output": "## Code Review Summary\n\n### Critical Issues\n..."
  }
}

​

Delegation Lifecycle

pending → running → completed
                        │
                  [quality check]
                        │
           ┌── pass ────┴──── fail ──┐
           │                         │
       accepted              revision_requested
                                      │
                                   running (retry)
                                      │
                               completed (or failed)

​

Team Memory

​

Example: Multi-Step Research Report

• System prompt: “You are a research director. Break complex research tasks into focused subtasks and synthesise professional reports.”

1. Ava (Researcher) — “Search and gather factual information, statistics, and expert sources on the given topic.”
2. Sam (Writer) — “Transform research notes into clear, structured prose. Use professional tone and logical flow.”
3. Smith (Editor) — “Review and polish written content. Fix grammar, improve clarity, ensure factual accuracy.”

1. Brain delegates “Research: AI adoption in healthcare 2025” → Ava
2. Brain delegates “Write a 1500-word report from these notes” + Ava’s output → Sam
3. Brain evaluates Sam’s draft — requests revision if below quality threshold
4. Brain delegates “Final editorial review” → Smith
5. Brain calls final_answer with synthesised report

​

Visual Workflow Builder (Canvas)

​

Canvas Features

• Drag agents from the sidebar onto the canvas to add them as workers
• Connect nodes by dragging edges to define delegation relationships
• Right-click context menu — Delete, Auto-layout, Set as Brain, Fit View
• Glassmorphism styling — frosted glass nodes with hover lift effects
• Searchable sidebar — filter agents by name or model

​

Live Execution Visualization

• Node states — Idle, Running (blue pulse), Completed (green ✓), Failed (red ✕)
• Camera auto-follow — Canvas pans to the currently executing agent
• Execution timeline — Step-by-step progress rail with clickable steps
• Transcript panel (T) — Real-time brain-to-worker message feed with delegation/result filters
• Tool heatmap — Tool usage stats with success rates

​

Keyboard Shortcuts

​

Auto-Layout

​

Position Persistence

​

Monitoring

• Delegations panel — each delegation with status, worker name, instruction, output, and revision count
• Messages feed — real-time log of brain decisions and worker responses
• Token usage — per-delegation breakdown
• Delegation depth — current iteration count in the orchestrator loop

​

Tips

• Brain model matters. The brain needs to reliably output JSON tool calls. Claude Sonnet 4+ and GPT-4o handle this well. Smaller or older models may produce malformed tool calls.
• Specific worker prompts = better delegation. The brain picks workers based on their name and description. Clear, focused descriptions (e.g. “TypeScript code reviewer” vs “AI assistant”) lead to better routing.
• Set quality threshold thoughtfully. Too high (0.9+) and the brain will loop excessively. Too low (0.3) and poor outputs get accepted. 0.6–0.8 is a good starting range.
• Watch delegation depth. If runs are looping on revisions, consider increasing max_delegation_depth or lowering the quality threshold, or improving the worker’s system prompt.
• Use team memory. After a few runs on similar tasks, team memory kicks in and the brain starts with better context.

​

Related

• Pipeline Teams — Fixed sequential steps, no dynamic routing
• Collaboration Teams — Agents discuss and reach consensus