Jev Decision Models Ai Agents Cline
- Source
- cline.bot
- Author
- Cline evaluation and engineering blog
- Date
Context
Cline, the AI coding , is using TypeSafe AI's Jev to make fast, narrow decisions inside its agent while a regular handles the language work. Jev is a decision model: instead of writing text, it takes a description of a situation, called a state, plus a set of questions, and returns for each question a pick from fixed options, a score on a scale, or a noul, the probability that a statement is true. In its blog post, Cline repeats the report that Jev is about 194 times faster and 445 times cheaper than traditional LLMs on TypeSafe's own workflow evaluations; Cline does not present its own measurement of those figures.
The clearest example is a check on a command an agent wants to run, an npm install of a package. Jev answered in 318 milliseconds, classifying the command as a network and system change with 82% probability, placing its risk in the middle band of three, worth a look, and giving an 82% probability that it downloads and runs code from the internet. A that runs before each tool call can use answers like these to block or pass the call on the spot.
Cline has built three plugins on this split. The jev-browser plugin gives Jev a list of the buttons, links and fields on a page and lets it pick each next action, after which Cline checks the final screenshot to confirm the job was done. The jev-playground plugin uses Cline to turn a vague request into a structured Jev request that a person can edit and run directly, with no LLM in the loop. The jev-guard plugin judges tool activity in a protected folder against the user describes in plain language, such as attempts to send refunds to other bank accounts through a support inbox agent.
The post shows a design pattern and one timed response, not a benchmark of its own. In the example, Jev's risk answer came with a confidence of only 0.58, so a guard built this way still needs a rule for uncertain cases. For testing of Jev outside TypeSafe's numbers, see this independent evaluation of Jev.
- Jev, from TypeSafe AI, does not generate text. It takes a structured state plus questions and returns a choice, a score, or a 'noul' (the probability a statement is true) for each, which an app can use to rerank or decide right away.
- In Cline's example, Jev judged an npm install command in 318 ms, put 82% probability on it downloading and running remote code, and classed it as a network and system change: the kind of check a before-tool hook can use to block or pass a call.
- Cline's jev-browser plugin gives Jev a list of actionable page elements (buttons, links, fields) and has it pick the next action, instead of sending a screenshot to a model at every step. Cline then checks a final screenshot to confirm the task was done.
- jev-guard screens every tool result inside a guarded folder through a Cline hook, so the agent does not have to remember to call the check. Cline turns a plain-language description of what to protect into guardrails with sample triggers to test.
- The split Cline describes: the LLM handles language and setup, deciding which facts and questions matter, while Jev handles the repeated quick decisions. The 194x faster and 445x cheaper figures are TypeSafe's own, from its workflow evals.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
- agent hook — A script the harness runs automatically at a fixed point in an agent's loop — before a tool call, after an edit, on session start.
- guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
Checking sign-in…
Loading comments…