It is easy to read “model for software” and assume Jev can replace the language model behind a coding assistant. TypeSafe’s coding-agent guide says otherwise: Jev does not write code, chat, or generate text. A coding agent can help you create an application that calls Jev, but Jev itself is a component for structured judgments inside that application.
A concrete division of work
Imagine an agent product that receives a user request. Its language model can plan, explain, and draft a response. Before choosing a tool, your application may need to route the request to one of a known set of handlers. Jev can evaluate that bounded decision, return a choice and probabilities, and let your code enforce the tool permissions and review rule. The agent still needs its normal language model for conversation and code generation.
This separation is useful because the route is testable. You can define valid options, collect a sample of user requests, compare selections to expected handlers, and inspect uncertain cases. The final action should remain behind ordinary checks: an apparently confident route is not permission to run a tool or reveal data.
Three patterns builders have published
Guard ambiguous tool calls
Jev-kit for Claude Code places a guard before tool use. The author says plain-code checks let about 93% of calls through in roughly 33 ms and send the gray cases to Jev. The model can judge whether a command resembles a risky read or broad search; hard permissions still belong in code. This is a useful split to copy conceptually: do not pay for a model decision when an exact rule already answers the question. The timing and pass-through rate belong to that author and setup, not to all coding agents.
Check edits against project rules
Enforce CLAUDE.md With Jev turns rules in a project instruction file into typed checks over replies and edits. Its author reports 93.3% precision and recall on 59 labeled examples across 19 rules, 348 ms p50, and about $0.045 per 1,000 checks. Those figures are a small, author-run evaluation, not an independent benchmark. The author also notes that exact-character and dataflow rules are weak spots. If a rule can be checked exactly by a linter, type checker, test, or permission boundary, use that instead.
Gate what an agent remembers
Jev Gates Agent Memory Saves compares a proposed memory with the original conversation. A tentative statement such as “we might use Postgres” should not become “we chose Postgres” in durable memory. On 100 synthetic cases at a .40 cutoff, the author says Jev kept 39 of 50 memories labeled worth keeping and saved none of the candidates labeled skip or defer. The median gate latency reported was 250 ms including network time. This suggests a pattern to evaluate, not a threshold to reuse: test on your own conversations and measure both false saves and useful memories lost.
Evaluate the full agent loop
Before adding a judge call, name the exact action it can influence: allow a tool, request confirmation, reject an edit, or save a memory. Gather examples with expected outcomes and include failures such as truncated context, ambiguous instructions, and adversarial text. Compare the model-assisted path with an exact-rule baseline and measure the cost of false approvals, not just average latency. Log enough context to explain a rejected action without copying secrets into logs. Keep a safe fallback for API failures and let a person review consequential disagreements.
How to begin
Pick one existing branch in your agent workflow. Write the states and allowed destinations on paper. Try a Choice question in the Playground, including requests that do not fit any destination. Add an explicit review or fallback path if needed. Only after the decision holds up on examples should you call the API from the product.
The official agent skill can give a coding agent context about TypeSafe’s API while it writes integration code. For the underlying design, read Choice, Score, and Noul and confidence and review.
Published 24 September 2026. The product distinction was checked against TypeSafe’s coding-agent documentation on that date.