Sub-Agents

What This Demo Shows

Multi-agent delegation with a live log. A supervisor LLM orchestrates three specialized sub-agents exposed as tools, and every delegation streams into the UI in real time via shared agent state.

• Three specialized sub-agents: research_agent (gathers facts), writing_agent (drafts prose), critique_agent (reviews drafts) — each is its own full create_agent(...)
• Sub-agents-as-tools: the supervisor calls them through @tool wrappers; each wrapper runs the sub-agent and appends an entry to the shared delegations state slot
• Live delegation log: the left pane renders delegations from agent state, growing as the supervisor fans work out

How to Interact

Click a suggestion chip, or type your own prompt. For example:

• "Produce a short blog post about the benefits of cold exposure training. Research first, then write, then critique."
• "Explain how large language models handle tool calling. Research, write a paragraph, then critique."
• "Summarize the current state of reusable rockets in 1 polished paragraph, with research and critique."

Watch the delegation log fill in as the supervisor runs research → write → critique.

Technical Details

• Each delegation tool returns a LangGraph Command(update={"delegations": [...], "messages": [ToolMessage(...)]}) — appending to shared state AND feeding the result back as a tool message in one step
• AgentState subclasses langchain.agents.AgentState and adds delegations: list[Delegation]; create_agent(..., state_schema=AgentState) makes it available on the frontend
• Frontend uses useAgent({ agentId: "subagents", updates: [UseAgentUpdate.OnStateChanged, UseAgentUpdate.OnRunStatusChanged] }) and reads agent.state.delegations + agent.isRunning to drive the log
• CopilotKit provider uses agent="subagents", backed by src/agents/subagents.py