Jev in Practice: Browser Use and Agent Workflows

Inspired by TypeSafe AI's Jev launch, Databricks built an enterprise API that turns raw text into structured decisions on governed data, callable in batch or real time.

The author built GTM bots on eight setups including Claude, Codex, Kimi, GPT-6, Grok and Jev, and shares five lessons before building and five on running the stack, centered on letting models judge while code handles permissions.

A use-case page from One shows how Jev and Mastra are combined to route inbound email in a shared inbox.

A configuration that turns Claude Opus 5.5 into an operating system for AI agents, splitting work between planner, researcher and builder roles, with DOTS and Jev as the decision layer.

A demo shows routing relevance, worker choice, retries, and tool-call safety to Jev as pick, score, or yes/no probability outputs instead of asking an LLM.

Dub first built its AI emoji picker with Jev, but feeding the full emoji catalog consumed many input tokens, so it refactored to a simple Upstash vector index and costs became nearly negligible.

The author calls DOTS + Jev the best agent team they have used and shares a full prompt that has Codex inspect and reuse existing dots-agent-team, Model Router, and Jev adapters.

Opus 5.5 plans, Sonnet 5.5 runs parallel work, Fable 5.1 advises, and Jev handles forks that need no thinking in under 12ms, cutting token use by 79%.

A Codex setup keeps GPT-6 Astra as a read-only advisor that speaks only before planning, when the same error repeats, and before declaring done, while GPT-6 Luna handles routine work and GPT-6.1 Sol plans and ships.

CMU proposes using JEV's calibrated probabilities as an LLM judge that accepts confident verdicts and escalates uncertain ones to GPT-6, saving 59% compute while being more accurate.

On the same 16-second sand video, Jev reads wind and track state once while Sonnet only handles targets still moving, avoiding repeated redraws of dead segments.

Someone built a tool that uses Jev to judge which parts of past chats are relevant to the current question, so you don't have to import your entire chat history into context.