Oct 14, 2026 19 min read

Durable Agents with LangGraph and LangSmith

Part 6: State that survives crashes, approvals that expire, evaluations that gate releases, and an authenticated streaming API.

👉

Recap

Part 5 turned the investigation into a team.

A deterministic intake routes the question. Three specialist workers with read-only tools run in parallel. An independent reviewer reconciles their findings. The one write action waits for the reviewer and a human to approve it.

Part 5 ended with three gaps:

Every checkpoint lives in memory. A process restart loses every paused approval and every run in flight.
An approval waits forever. There is nothing that expires it.
Lastly, nothing catches prompt regression. Suppose we deleted the reviewer's instruction to judge whether the evidence supports the proposed action. All 81 tests from Parts 1 to 5 would still pass, since they use scripted models, which do not read prompts, so no test could notice that.

This part closes these gaps. State moves to Postgres and survives a crash. Approvals expire by policy, and every decision goes into an audit table. Evaluations check the nodes that a prompt change can break, and a gate fails the build when their scores fall below the thresholds. The team also gets an authenticated API that streams its progress and pauses for approval.

Prerequisites: Parts 1 to 5 of this series. You can read them below: