Human in the Loop workflows and/or Generative UIs support

[pcuci]

re: https://github.com/langchain-ai/agent-inbox

this might be totally out of scope for chat.md, but was curious how you'd approach this kind of user interaction?

when an assistant needs to, in its own workflow, ask for additional information from the user (I think that would map to an interrupt inside a tool call in chat.md speak), how might the human provide the additional information/feedback?

there is a nascent generative ui trend, and I'm wondering if it could be supported by chat.md in the future(?)

[rusiaaman]

This is a cool agentic feature.

Currently chat.md isn't an autonomous agent. It's more like a chat application that has some tool calling using MCPs.

To have chat.md have such agentic features, I'd bet on implementing A2A protocol from google, which works well with MCP.

So chat.md could be both an agent implementing A2A protocol and an MCP client.

For obvious reasons, it'll have to be interaction over files not full GUI. So any multimedia will be a file linked in the markdown. Any interaction will be using file edit operations.

What do you thik about that?

[pcuci]

Looks like at least one person is working to turn LangGraph agents into MCP tools: https://github.com/Adish03101/langstuff/blob/master/env/Lib/site-packages/langgraph_api/api/mcp.py

My guess is this will land in LangGraph once Streamable HTTP deprecates Server-Sent Events, at the protocol level: https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/docs/specification/draft/basic/transports.mdx#streamable-http

Not clear if we would still need LangGraph agents to act as MCP servers should chat.md natively support A2A: https://github.com/google/A2A/tree/main/samples/python/agents/langgraph#how-it-works