Cross-Agent Collaboration
Some tasks are too complex for one agent to handle. In these cases, multiple agents can divide the work and collaborate, each leveraging their expertise.
When Multi-Agent Collaboration is Needed
| Scenario | Agents Involved | Collaboration Method |
|---|---|---|
| Review contract + Generate report | Legal Advisor + Document Assistant | Legal advisor reviews, document assistant formats report |
| Analyze data + Write weekly report | Data Analyst + Work Assistant | Analyst processes data, assistant writes into weekly report |
| Translate document + Professional proofreading | Translation Assistant + Domain Expert | Translation assistant does initial translation, expert proofreads terminology |
| Competitive analysis + Strategy recommendations | Market Analysis + Strategy Consultant | Analyst collects data, consultant gives strategy |
How to Trigger Collaboration
Method 1: You Actively Arrange
You can directly specify which agents participate:
You: "Have the legal advisor review this contract first, then hand the results to the document assistant,
to generate a formal review report."
Method 2: Agent Automatically Identifies
When the agent finds it needs other expertise during task execution, it proactively suggests collaboration:
Agent (Legal Advisor): "I have completed the legal review of the contract, but found that the contract
involves international trade clauses. This is beyond my expertise.
Suggestion: Have the trade advisor assist in reviewing clauses 15-18 related to international trade.
Allow?"
Information Transfer in Collaboration
When multiple agents collaborate, information transfers between them in an orderly manner:
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Legal Advisor│ ───→ │ Trade Advisor│ ───→ │ Document Assistant│
│ │ │ │ │ │
│ Review contract clauses │ │ Review trade clauses │ │ Generate formal report │
│ │ │ │ │ │
│ Output: │ │ Output: │ │ Output: │
│ - Penalty clause risks │ │ - Trade clause analysis │ │ - Complete review report │
│ - Payment term issues │ │ - Compliance suggestions │ │ │
└──────────────┘ └──────────────┘ └──────────────┘
Content Transferred
Structured task handoffs can include the following information:
| Transfer Item | Explanation |
|---|---|
| Task description | What needs to be done by the other party |
| Input data | Files or information the other party needs to process |
| Completed work | Output from previous agents |
| Constraints | Rules or requirements to follow |
| Context | Relevant background information |
For lighter-weight collaboration subtasks, the delegating agent also decides on its own how much background the collaborator actually needs. If the work is something like confirming a conclusion or checking a number—things that don't require reviewing the full history—it can have the collaborator skip the full memory retrieval and carry only the information required to finish that specific task. This lets the collaborator respond faster and use fewer resources. This judgment is made by the agent based on how substantial the task is; it isn't something you configure manually, and you'll only notice it as a note like "background information trimmed" in the collaboration receipt or process log.
Checking delegation progress
Expand a delegation card to inspect each participant's state.
| Card label | Meaning |
|---|---|
| Sequential delegation | Work progresses in order |
| Parallel delegation | Work can proceed concurrently; results may arrive at different times |
The label describes the delegation strategy. Participant states and execution records show whether work has started or finished. After reopening a conversation, inspect those records to avoid duplicating a task that is still running.
To delegate board work, register the task, assign its owner, then ask that owner to execute it. Creating a task or sending a message alone does not start execution. See Long-running tasks.
Letting one agent update another agent's configuration
When you explicitly request it, one agent can update another agent's AgentFS files, such as persona.md and principles.md. A team lead adjusting a member's role is the most common case, but team membership does not silently grant background access:
- The change must come from your explicit request or consent; an agent must not rewrite another agent's persona or principles on its own
- Cross-agent writes display an approval card and are not saved if you reject them
- Another agent's memory does not become approval-free merely because both agents belong to the same team
- Changes are recorded in tool audit data and Git history and can be reviewed or reverted through Rewind checkpoints
- Afterward, the editing agent should identify the target agent, affected files, and main changes
Sensitive paths such as .env, .ssh, credential files, and Git internals remain blocked. User approval does not bypass these hard protections.
Monitoring Collaboration Progress
When multiple agents collaborate, you can see the overall progress panel:
┌──────────────────────────────────────────────────────┐
│ 🤝 Collaboration Task: Comprehensive Contract Review │
├──────────────────────────────────────────────────────┤
│ │
│ Participating agents: 3 │
│ Overall progress: ━━━━━━━━━━━━━━░░░░░░ 75% │
│ │
│ ✅ Legal Advisor Complete │
│ └─ Reviewed 15 clauses, found 2 risk points │
│ │
│ 🔄 Trade Advisor In progress │
│ └─ Analyzing international trade clauses... │
│ │
│ ⏳ Document Assistant Waiting │
│ └─ Waiting for previous tasks to complete before starting │
│ │
│ ┌──────────┐ ┌──────────┐ │
│ │ View Details │ │ ⏸ Pause │ │
│ └──────────┘ └──────────┘ │
│ │
└──────────────────────────────────────────────────────┘
View Detailed Progress of Each Agent
Click on an agent to view its detailed execution status, same as the single-agent execution monitoring interface.
Intervening in Collaboration Process
You can intervene at any time during collaboration:
You: "The trade advisor's analysis doesn't need to be too detailed, just confirm whether
export control risks exist, skip the rest."
Trade Advisor: "Received, I will focus on reviewing export control related clauses."
Graph and Table Views in the Decision Tree Workbench
In the Decision Tree Workbench, use the view selector to switch between Tree, Board, Graph, and Table. Graph displays the nodes, conditional edges, and branch topology in the decision definition. Table lists the properties of variables, input questions, nodes, and edges without requiring a canvas layout. Selecting an entity in either view selects its definition for inspection in the properties area. Graph and Table are for browsing and selection; edit and save the decision definition in the workbench’s structure or properties areas.
Answering questions and confirming statements
The decision workspace lists missing inputs separately from business statements awaiting confirmation.
- Answer the questions first. Mark uncertain answers as unknown; answering does not confirm a statement.
- If the conversation entry shows pending confirmations, use its confirmation action to review the pending statements.
- Automatic continuation waits until all required statements in the current workspace are confirmed.
After a submission timeout, inspect the saved result before retrying. An unconfirmed draft is not a submitted result.
Collaboration Receipts
After collaboration tasks complete, a comprehensive receipt is generated, containing contributions from all participating agents:
│ 📋 Collaboration Receipt │
│ │
│ Participants and contributions: │
│ │
│ 1. Legal Advisor │
│ - Clauses reviewed: 15 │
│ - Risks found: 2 │
│ - Time: 3 minutes │
│ │
│ 2. Trade Advisor │
│ - Clauses reviewed: 4 │
│ - Compliance confirmation: No export control risks found │
│ - Time: 1.5 minutes │
│ │
│ 3. Document Assistant │
│ - Report generated: 1 (2.3KB) │
│ - Time: 2 minutes │
│ │
│ Total time: 6.5 minutes │
│ If executed sequentially (non-collaboration): ~8 minutes │
Not all tasks need multi-agent collaboration. Simple tasks are more efficient with one agent. Collaboration is only valuable when tasks involve multiple professional domains or require parallel processing of multiple subtasks.
Congratulations on completing all the learning for "Delegating Tasks"! You now know:
- How to Give Tasks
- How to Confirm Plans
- How to Monitor Execution
- What is Human Gate
- How to interpret Execution Receipts
- How to use Cross-Agent Collaboration