Deploy deep agents to LangSmith

[GeneralJerel]

Overview

Deploy the LangGraph deep agent (apps/agent) to LangSmith's managed cloud platform (formerly LangGraph Platform) for production hosting, observability, and tracing.

Background

The agent already uses the deepagents SDK (create_deep_agent()) built on LangGraph, and has a langgraph.json config in place. LangSmith Cloud provides managed deployment with built-in tracing, streaming, and persistence — replacing the current self-hosted FastAPI/Uvicorn setup.

Current State

• Agent framework: LangGraph 1.0.7 + DeepAgents via create_deep_agent()
• Entry point: apps/agent/main.py → exports graph object
• Config: apps/agent/langgraph.json already exists with graph sample_agent pointing to ./main.py:graph
• Dependencies: managed via uv + pyproject.toml
• Hosting: currently self-hosted via uvicorn on port 8123
• Checkpointing: in-memory BoundedMemorySaver (200 threads, no persistence across restarts)

What Needs to Happen

1. LangSmith Account & Deployment Setup

• Ensure LangSmith Plus plan (or higher) is active — required for cloud deployments
• Connect the GitHub repo to LangSmith via Deployments → + New Deployment

2. Configuration Updates

• Review langgraph.json — the env path currently points to ../../.env (relative to repo root), which may need adjustment for cloud deployment
• Add LANGCHAIN_API_KEY, LANGCHAIN_TRACING_V2=true, and LANGCHAIN_PROJECT to environment config
• Ensure OPENAI_API_KEY is set as a deployment secret in LangSmith
• Verify deepagents and all dependencies resolve correctly in the cloud build environment

3. Compatibility Considerations

• CopilotKitMiddleware: Verify it works within LangGraph Platform's execution model (not running inside FastAPI directly)
• BoundedMemorySaver: Cloud deployments provide managed persistence (Postgres-backed checkpointer) — evaluate whether to replace the custom in-memory saver
• Skills directory: Confirm skills/ markdown files are included in the deployment bundle (they're loaded via file path at runtime)
• ag-ui-langgraph endpoint: Determine if the FastAPI wrapper (add_langgraph_fastapi_endpoint) is still needed or if LangSmith handles this natively

4. Frontend Integration

• Update the frontend CopilotKit API route to point to the LangSmith deployment URL instead of localhost:8123
• Add deployment URL as an environment variable for the Next.js app
• Test bidirectional agent state sync (todos) through the cloud-hosted agent

5. Observability

• Confirm traces appear in LangSmith for all agent runs
• Set up a LangSmith project for organizing traces
• Verify tool call traces (query_data, plan_visualization, manage_todos, etc.) are captured

Risks & Open Questions

• Cold start latency: LangSmith cloud deployments may have cold starts — measure impact on UX
• Cost: Plus plan is $39/seat/mo + usage-based billing for agent runs and uptime
• Skills file loading: The agent loads skills from Path(__file__).parent / "skills" — need to verify this resolves correctly in the containerized cloud environment
• Rate limiting: Current app has optional IP-based rate limiting in FastAPI middleware — this won't carry over to LangSmith; determine if platform-level rate limiting is sufficient

References

• LangSmith Deployment Docs
• Deep Agents SDK
• LangGraph Cloud Application Structure