Jev in Practice: Browser Use and Agent Workflows

Jev founder Diogo Amogo wrote a PDF on building a Jev harness, claiming it is 200x faster and 400x cheaper, and says you can hand the PDF to Claude Code or Codex to rebuild its own setup.

A rundown of September releases singles out TypeSafe AI's Jev: no chat window, just choices or probabilities your code can branch on, plus its price and speed claims.

The author uses jev-opus to hand fork decisions like effort, tools, retries and escalation to Jev, leaving Opus 5.5 to write and plan, with per-message statements preserving the prompt cache.

Emails pass through Jev for confidence routing, and only the uncertain 10% escalate to Kimi K3, bringing a 10,000-email monthly bill from $300 to about $9.42, or $27.42 at a 30% escalation rate.

The author uses Jev as an LLM judge to run agent evals, calling it the most cost-effective and fastest approach, and has integrated it into his own intelligence harness.

An eight-point comparison of positioning, input, output, latency, accuracy, price, specs and availability concludes that Jev is the first pick for high-frequency text routing.

SGLang now lets a plain LLM or VLM act as a classification and scoring model, and adds a Jev-compatible /v1/systemone endpoint so the Noul, Choice and Score interfaces can run on open models.

A Thai user rounds up more than 20 DevDay announcements, noting that item four, the Decisions API with computer use, is Jev, and that GPT-6.1 Sol costs about 10.5x less per task than Sonnet 5.5.

jevgrep asks whether session expiry is handled in each folder, file and function, then passes only the surviving lines to agents like Claude Code and Codex.

A Jev-style gate filters out 97% of crypto news noise, sends only ambiguous high-impact events to Claude, and code weights news at a 30% cap; a 22-round replay takes $100 to $2,981.

Someone mapped an entire dev team onto 160+ Claude Code subagents split across API design, backend, review, QA, security and DevOps, calling it Jev-style engineering: stop making one agent do the whole team's job.

Someone organized Jev's uses in agents from the TypeSafeAI docs into nine schemes: model routing, guardrails, tool-call gating, inbox triage, reranking, LLM evals, bulk labeling, real-time control and confidence gating.