LangChain: Building production agents with Jev and LangGraph

It is TypeSafe's philosophy as LangChain quotes it: code owns the workflow and AI handles narrow, structured decisions. The post says frontier models got treated like god for every task with fuzzy inputs, and god is expensive and slow, so you pay for that generality even when you needed a yes or no.

Structured answers come back as typed answers with probabilities, so code can branch predictably; parallel evaluation lets you ask many questions about the same state at once; decisions are cheap enough to make many in a run; and system one is designed to return stable answers across repeated evaluations.

In an early Jev-as-a-judge experiment cited by the post, Jev's scores barely moved across 100 repeated runs, far less than any LLM judge LangChain tested, which the post frames as a welcome change from the non-determinism of LLMs.

Litigation discovery: a company must review every page before producing documents, and a single matter can run to hundreds of thousands of pages. For each page Jev answers three questions in one request, whether the page is responsive, contains personal information, or might be privileged, and each answer maps to a route in the graph.

LangChain ran the same graph with Jev handling classification and with Sonnet as the judge, and says Jev was 5-6x faster on the classification step across trials. Both runs are traced in LangSmith, where dedicated views show decision-model inputs and calibrated outputs.

Browserbase rebuilt Stagehand's act() around the idea: Stagehand marks interactive elements, Jev picks the action type and best candidate element, and anything below a 0.7 confidence threshold falls back to an LLM. In early testing, median act() latency dropped from 1.97 seconds to 0.46 seconds, about 4.3x faster.

Jaya Gupta's phrase for capabilities being pulled apart and sent to the cheapest model that can handle each one. Jev pulls out judgment; a runtime then stitches the pieces back together by routing each step and deciding when to escalate, a shift from frontier by default and optimize later to cheap by default.