Agent Chain

[iceman0110]

Since the deprecation of Agent Chain, what is the recommended approach for users to create new workflows involving several existing agents?

For example, we need a defined sequence: Research → Draft → Review, where each agent passes its result to the next.
Can this currently be configured through the UI using Subagents or Handoffs while maintaining that sequence, or does it require API configuration?

We would appreciate guidance on the supported approach for users migrating from Agent Chains.

[IlyaKhalafi]

The fixed team needs the Agents API. UI Subagents and Handoffs are model-chosen delegations or transfers, so neither sets this order.

Create a parent Agent in Agent Builder, with Research, Draft and Review as separate saved members. Using the signed-in user's JWT, read GET /api/agents/<id>/expanded, then merge this into its existing subagents object and send that object through PATCH /api/agents/<id>. These are the Agent Builder routes, distinct from machine-management endpoints. The Agents endpoint's subagents capability must be enabled.

{
  "subagents": {
    "enabled": true,
    "allowSelf": false,
    "graphs": [{
      "type": "research_draft_review",
      "name": "Research, draft and review",
      "description": "Research a topic, draft the result and review it",
      "agent_ids": ["agent_research", "agent_draft", "agent_review"],
      "entry_agent_id": "agent_research",
      "result_agent_id": "agent_review",
      "edges": [
        {"from": "agent_research", "to": "agent_draft", "edgeType": "direct"},
        {"from": "agent_draft", "to": "agent_review", "edgeType": "direct"}
      ]
    }]
  }
}

Replace the placeholders with saved IDs. Every chat user must be able to access all members; otherwise the team can become unavailable for that user.

In SDK 4.0.4, these direct edges carry shared message history, including earlier results. Only Review's configured result returns to the parent, so its instructions should include everything the user needs.

The parent can skip or repeat the team call. Explicit delegation instructions can help, but this does not guarantee one run per message like an always-run chain.

This is based on LibreChat commit e1dfc10, the development docs and SDK 4.0.4 source. I have not run the complete workflow. Check that your deployment includes saved-team support.