Jev in Practice: Browser Use and Agent Workflows

Swapping the model inside Reticle's harness to Jev lowered cost from $2.12 to $0.006 and tokens from 8,048,739 to 161,331, with both finding the same 100x refund bug.

The author credits Jev's speed and cost but notes it is only a classifier limited to preset labels that can still pick wrong, and that the RLCD method and benchmark are unpublished.

Jev replaces a full LLM call with a typed choice, score, or yes/no to decide an agent's next tool, and is claimed to be 193x faster and 444x cheaper than a frontier LLM on routing tasks.

A developer argues the counterintuitive setup: rather than having an LLM agent orchestrate Jev calls, deterministic workflows should lead, using Jev for judgment and to decide when to delegate to an LLM.

A developer has their agent integrate Jev across the whole app and run QA on performance and cost comparisons, saying every part of Gab AI will get faster and cheaper.

A developer's experiment finds that optimizing instructions and state representation, as in dspy.Flex or GEPA, also improves Jev's performance noticeably.

A post asks whether anyone is trying Jev for scalable oversight, suggesting it could help determine if a model is behaving in an aligned way.

Codex's live output, Git changes and AGENTS.md are combined into one observation, and Jev scores the probability that the worker is drifting from AGENTS.md, steering the active turn past a threshold and stopping it if drift continues; it is now part of Foreman.

An eight-step plan that splits deterministic math from Jev's judgment, runs six parallel decision questions with confidence gating and Kelly sizing, and includes backtesting and calibration checks.

The author uses Jev to filter their todo list and surface items like revenue increases, getting stuck, or switching SaaS providers that keyword search would miss.

Argues for a dedicated decision layer between LLM reasoning and tool execution, reporting up to 193x faster and 444x cheaper results, with a state-to-decision-to-action-to-verification blueprint.

Using spam classification as an example, Jev takes an input and an output schema and returns numbers like 0.90 plus a category from a fixed list, with no reasoning text, suited to fast low-cost binary decisions in code.