LangChain and Deep Agents (retired)

Fountain used to fit into LangChain and Deep Agents as a subagent: a FountainAgent runnable in one Python file, talking to Fountain over the OpenAI-compatible API as a ChatOpenAI model, with a LangGraph thread_id mapped to a sandbox. That API is gone, removed in the release that carries ADR 0057, and the integration and its examples/deepagents-contractor example went with it.

This page stays because the URL is in other people's notes.

What has no replacement

Be clear about the size of this, because it is the biggest thing the retirement costs:

• A stock ChatOpenAI pointed at Fountain. The integration worked because Fountain spoke a dialect LangChain already had a client for. It does not any more, so there is no base URL to swap in.
• The role: "tool" continuation loop. LangChain's tool calling returned results to Fountain as chat messages, and the request-defined tool bridge turned those into answers for the agent. Both halves are retired (ADR 0057). Fountain's native API has no equivalent: an agent's tools are configured on the agent, not defined per request by the caller.

Nothing in the native API gives either of those back, and this page is not going to pretend otherwise.

What you can still do

Drive Fountain from Python directly, with the Python SDK or the conversation API: create a conversation for the work, prompt it, follow its stream, read the result, and hand that back to your orchestrator yourself. That replaces the delegation — an orchestrator handing work to a Fountain agent and reading its report — which is what most people used this for. It is a wrapper you write, not a model object LangChain already understands.

If the Fountain agent needs to call back into your application mid-run, configure that as an MCP server on the agent, with ${VAR} references resolved from the environment and the vault. That path is fully supported and is not the retired bridge. Read Plug into Fountain.