LangGraph Testing in CI/CD

LangGraph gives you stateful agent flows with graph-based orchestration, but you still need a reliable way to verify graph behavior after prompt, model, tool, or node changes. EvalView provides that regression loop with a dedicated LangGraph adapter featuring auto-detection, streaming support, and native thread tracking.

What matters for LangGraph agents

LangGraph agents have unique testing needs beyond simple input/output checks:

• state-driven multi-turn flows — does the graph handle conversation state correctly?
• tool-node selection and sequence — does the agent route to the right nodes?
• branching behavior and fallbacks — do conditional edges fire correctly?
• user-visible output regressions — does the final response degrade?
• cost and latency changes by run path — does a specific graph path cost more after changes?

Quick start with LangGraph

EvalView auto-detects LangGraph agents. Start your agent server, then connect and test:

# Start your LangGraph agent
uvicorn main:app --reload --port 8000

# Connect EvalView (auto-detects LangGraph)
evalview connect --endpoint http://localhost:8000/api/chat

# Generate tests from the running agent
evalview generate

# Snapshot approved behavior
evalview snapshot tests/generated --approve-generated

# Check for regressions
evalview check tests/generated

CI integration for LangGraph

Add regression checks to your CI pipeline so every PR is validated:

- name: Check for regressions
  uses: hidai25/eval-view@v0.6.0
  with:
    openai-api-key: ${{ secrets.OPENAI_API_KEY }}
    fail-on: REGRESSION

Multi-turn test cases for LangGraph

Multi-turn tests are especially useful for LangGraph because they verify state management across conversation turns:

name: refund-needs-order-number
turns:
  - query: "I want a refund"
    expected:
      output:
        contains: ["order number"]
  - query: "Order 4812"
    expected:
      tools: ["lookup_order", "check_policy"]
      forbidden_tools: ["delete_order"]
      output:
        contains: ["refund", "processed"]

Troubleshooting

• "Cannot connect to agent" — verify your server is running with curl and check the endpoint path
• "Wrong endpoint" — update .evalview/config.yaml with the correct URL
• "Tool names don't match" — run with --verbose to see actual API responses, then update YAML files

CI/CD Integration Guide | Regression Testing Guide | Back to EvalView homepage