feat : Add deep agent support

[gauravsingh-rajput]

Add deep agent support

Adds a thin wrapper around deepagents.create_deep_agent to expose deep agents through uipath_langchain.

Changes

• src/uipath_langchain/agent/deep/ — New module with create_deep_agent wrapper that accepts Sequence types for tools/subagents and converts them to lists before forwarding to the upstream
• tests/agent/deep/test_create_deep_agent.py — 4 tests covering: custom tool registration, built-in tool presence (write_todos), FilesystemBackend support, and tuple-to-list conversion
• pyproject.toml — Added deepagents dependency

[cristipufu]

@gauravsingh-rajput What's the point in adding this dependency in uipath-langchain for a method wrapper?

[gauravsingh-rajput]

@gauravsingh-rajput What's the point in adding this dependency in uipath-langchain for a method wrapper?

@cristipufu this is intentionally incremental — right now it’s a thin wrapper, but the goal is for uipath-langchain to own agent construction and the deepagents dependency (same pattern as create_agent for react agents).

My understanding is that both coded agents and low-code agents (uipath-agents-python) go through this layer, which keeps deepagents as a transitive dependency and gives us a single place to add specific behavior (meta prompt, guardrails, custom backends). The switch between regular aka shallow agents and deep agents remains in uipath-agents-python.

In here we can add things like a custom backend which supports a sandboxed code execution and storage, and this abstraction ensures both paths benefit without changes in multiple places.

[cristipufu]

@gauravsingh-rajput The meta prompts live in uipath-python, and guardrails can be handled through middlewares since they aren't specific to deep agents.

The only reason we'd need to add the dependency here would be for custom backends/storage. But given that we run these on serverless with direct file system access, I'm not convinced we'll actually need custom implementations.

I gave this some thought while building out this sample, and my feeling is that we should hold off on adding the dependency for now. Let's keep experimenting and only pull it in once we have a clear need for it.