For the use of asynchronous sub-agents, the ASGI protocol must utilize langgraph.json. How can it be supported in the case of dynamic multi-agent systems

[yangzeyi666]

Feature description

For the use of asynchronous sub-agents, the ASGI protocol must utilize langgraph.json. How can it be supported in the case of dynamic multi-agent systems

Proposed solution (optional)

No response

Additional context (optional)

No response

[Hunter Lovell (hntrl)]

Converting to a discussion item since this is more of a usage question instead of a pointed feature request/ issue.

Could you elaborate more on what you mean by dynamic multi-agent systems?

[Gautam Vhavle (GautamVhavle)]

langgraph.json is how LangGraph Server / the ASGI app discovers graphs at process start. It is a static registry: path, id, env. It is not where you describe "which subagent exists this request."

For dynamic multi-agent, keep one (or a few) entries in langgraph.json, and let that graph construct or select subagents at runtime.

{
  "dependencies": ["."],
  "graphs": {
    "agent": "./src/agent.py:graph"
  }
}

# src/agent.py
from langgraph.graph import StateGraph
from langgraph.types import Send

def route(state):
    # decide which specialists to run for *this* request
    return [Send("specialist", {"task": t}) for t in state["tasks"]]

builder = StateGraph(...)
# compile once
graph = builder.compile()

Patterns that work with the ASGI server:

1. Single parent graph, dynamic Send / ToolNode. The compiled topology is fixed. Which specialist runs is data. This is the usual "dynamic multi-agent" on LangGraph Server.
2. Subagents as tools (create_deep_agent / SubAgentMiddleware). The parent is the graph in langgraph.json. Subagents are created in process when the parent tool is invoked. They do not each need their own langgraph.json entry.
3. Factory graph. Export graph as a compiled app whose nodes look up a registry (state["agent_id"] -> runnable). Still one ASGI graph. Do not generate a new langgraph.json per user.

What does not work: writing a new langgraph.json at request time and expecting the already-running ASGI server to pick it up. The server reads that file on boot.

Async subagents: they can be async nodes inside that one compiled graph. You do not need a second ASGI app. If you are hitting "must use langgraph.json" it is because the process you launched is LangGraph Server / langgraph up, which only mounts graphs listed there. Point it at the parent, keep dynamism inside the parent.

Happy to sketch a registry-based parent if you say whether the set of agents is per-tenant, per-user, or per-request. I have not seen your current langgraph.json.