{"slug": "langchain-adds-typesafe-s-jev-to-the-agent-control-loop", "title": "LangChain adds TypeSafe's Jev to the agent control loop", "summary": "LangChain integrated TypeSafe AI's Jev decision model into its agent control loops via a new `TypeSafeClassifier` interface, announced in a September 17 technical post, giving agent builders typed, confidence-bearing decisions for routing, escalation and tool-call checks. The integration followed TypeSafe's September 15 Jev launch and a $40M seed round led by DCVC that valued the San Francisco company at $200M, according to Forbes. TypeSafe was founded in 2024 by Diogo Almeida, Sasha Sheng and Erik Gafni; Jev accepts application state and returns choices, scores or yes-and-no probabilities rather than generating prose.", "body_md": "# LangChain adds TypeSafe's Jev to the agent control loop\n\n**The integration gives developers typed answers and confidence scores for routing, escalation and tool-call checks inside AI agent workflows.**\n\n        By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)\n        · Published \n\nPrimary source: [LangChain on X](https://x.com/LangChain/status/2100754885918830734)\n\n## Why it matters\n\nLangChain gives TypeSafe a direct path into agent infrastructure, where Jev's speed and confidence claims can be tested against ordinary structured-output model calls.\n\n[Diogo Almeida (@CompleteSkeptic)](https://x.com/CompleteSkeptic?ref=runtimewire), [Sasha Sheng (@hackgoofer)](https://x.com/hackgoofer?ref=runtimewire) and [Erik Gafni (@EGafni)](https://x.com/EGafni?ref=runtimewire) have landed an early distribution win for [TypeSafe AI](https://typesafe.ai/?ref=runtimewire): [LangChain](https://www.langchain.com/blog/building-a-harness-with-jev?ref=runtimewire) has integrated the founders' Jev decision model into the control loops developers use to build AI agents.\n\nIn a [September 17th technical post](https://www.langchain.com/blog/building-a-harness-with-jev?ref=runtimewire), LangChain said the integration gives agent builders typed, confidence-bearing decisions for routing and control.\n\nThe post exposes Jev through `TypeSafeClassifier`, allowing an application to submit its current state and a set of typed questions through `.invoke()` instead of asking a chat model to generate another response.\n\nLangChain's technical post followed [TypeSafe's September 15th launch post for Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev?ref=runtimewire) and [DCVC's September 15th announcement of a $40M seed round](https://www.dcvc.com/news-insights/typesafe-emerges-from-stealth-with-a-new-way-of-doing-ai/?ref=runtimewire), which the firm led. [Forbes reported](https://www.forbes.com/sites/the-prompt/2026/09/15/this-200-million-startup-wants-to-fix-ais-overconfidence-problem/?ref=runtimewire) that the financing valued the San Francisco company at $200M, citing a person familiar with the transaction.\n\nAlmeida started TypeSafe in 2024 after leaving OpenAI, where he co-authored the InstructGPT paper and contributed to the [GPT-4](https://runtimewire.com/models/openai/gpt-4) technical report. He had previously worked at Google Brain. Sheng was a research engineer at Meta and FAIR, while Gafni previously co-founded the DNA-sequencing AI company Ravel and worked at Invitae and Freenome, according to [TypeSafe's team page](https://typesafe.ai/team?ref=runtimewire).\n\nTheir founding thesis came from a problem Almeida said occupied him for four years. \"Models have been superhuman at chat for years, so where is all the automation?\" he wrote in [TypeSafe's introduction to Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev?ref=runtimewire). His answer is that models optimized for conversation are a poor default for the numerous bounded judgments software makes while completing a task.\n\n### A model for the decisions between model calls\n\nJev does not generate prose. It accepts application state, such as a support ticket, log entry, message history or structured record, and evaluates questions whose possible answers are defined in advance. Jev returns choices, scores or yes-and-no probabilities, along with confidence measures that code can use to determine its next step.\n\nLangChain's integration packages that behavior for agent workflows. A developer can install `langchain-typesafe`, set a TypeSafe API key and call `TypeSafeClassifier` from a node, middleware hook or tool. Multiple questions can be evaluated against the same state in one request, with TypeSafe saying Jev handles those questions in parallel.\n\nThe immediate uses are mundane by design. Jev can decide whether a customer request is urgent, choose which model should receive a task, flag an output for verification, escalate a case to a person or check whether a proposed tool call looks risky. LangChain's experimental `AutoModeMiddleware`, for example, uses Jev to classify tool calls before execution and can block actions involving tools such as a shell.\n\nThat gives TypeSafe a clearer position than another general-purpose model release. Jev is being inserted between the larger model calls, where an agent repeatedly decides what to do next. Almeida is betting that a specialized model can perform those narrow judgments faster and more predictably while leaving open-ended reasoning and text generation to conventional language models.\n\nLangChain makes that separation explicit. Its post says Jev is not a drop-in replacement for an LLM. The proposed architecture keeps a larger model at the center of the agent and assigns Jev the repeated classification work surrounding it.\n\n### TypeSafe's speed claims still need production evidence\n\nTypeSafe prices Jev at $0.042 per million input tokens with output free, while [Vercel's AI Gateway lists it](https://vercel.com/ai-gateway/models/jev?ref=runtimewire) at about $0.04 per million input tokens. TypeSafe claims response latency of 70 to 500 milliseconds and says Jev was 193.6 times faster than frontier language models in its workflow benchmarks. LangChain summarized the upper end of TypeSafe's claims as 200 times faster inference and 400 times lower cost on classification tasks. [TypeSafe's technical post](https://typesafe.ai/blog/introducing-system-one-models-and-jev?ref=runtimewire) contains the company's pricing and benchmark disclosures.\n\nThose figures come from TypeSafe's own evaluations. The company says its published speed tests generally ran from laptops on the West Coast, where its service is based. It also acknowledges that one demonstration used a short, information-dense input that favored Jev's parallel sampling design.\n\nThe workflow evaluation has a larger limitation: TypeSafe compared model outputs with the average probabilities produced by large external models, rather than against independently established ground truth. TypeSafe also says the workflows were created by members of its model capabilities team, leaving room for selection or design bias even if the examples were outside Jev's training data.\n\nTypeSafe's claim of zero type errors is narrower than a claim of zero bad decisions. Jev cannot return arbitrary prose outside the output schema because its architecture does not generate strings. A perfectly typed answer can still classify the underlying situation incorrectly, and developers will need to test whether Jev's confidence scores track real accuracy on their own workloads.\n\nThat distinction matters most in the use cases LangChain is opening up. A routing error may waste money by sending a simple task to an oversized model. A mistaken escalation decision can bury an urgent support case. A faulty safety classification can approve the tool call the control layer was meant to stop.\n\n### LangChain gives the thesis somewhere to run\n\nTypeSafe spent two years building Jev in stealth. LangChain now gives the model a route into existing agent code before TypeSafe has established public evidence of customer-scale usage.\n\nThe integration also puts TypeSafe's argument into a form developers can test without redesigning an entire application around a new model provider. Jev can be installed as a classifier beside an existing agent, assigned one routing or verification task and measured against the structured-output call it replaces. That lowers the adoption cost and makes the comparison harder to hide behind a benchmark.\n\nFor Almeida, Sheng and Gafni, LangChain's support is an early test of whether their model category is useful outside TypeSafe's demos. The $40M seed round financed a broad claim about machine-native intelligence. The LangChain integration reduces that claim to a practical question: does an agent work faster, cost less and fail less often when a specialist makes its routine decisions?", "url": "https://wpnews.pro/news/langchain-adds-typesafe-s-jev-to-the-agent-control-loop", "canonical_source": "https://runtimewire.com/article/langchain-adds-jev-decision-model-agent-workflows", "published_at": "2026-09-18 05:29:30+00:00", "updated_at": "2026-09-18 05:54:17.766214+00:00", "lang": "en", "topics": ["ai-agents", "ai-products", "ai-startups", "developer-tools", "ai-tools"], "entities": ["LangChain", "TypeSafe AI", "Jev", "TypeSafeClassifier", "Diogo Almeida", "Sasha Sheng", "Erik Gafni", "DCVC"], "alternates": {"html": "https://wpnews.pro/news/langchain-adds-typesafe-s-jev-to-the-agent-control-loop", "markdown": "https://wpnews.pro/news/langchain-adds-typesafe-s-jev-to-the-agent-control-loop.md", "text": "https://wpnews.pro/news/langchain-adds-typesafe-s-jev-to-the-agent-control-loop.txt", "jsonld": "https://wpnews.pro/news/langchain-adds-typesafe-s-jev-to-the-agent-control-loop.jsonld"}}