Jev in Practice: Browser Use and Agent Workflows

An article uses a support-agent trace to show that 37 of 44 model calls are decisions and 7 are writes, and suggests returning typed judgments from Jev while leaving the writing to another model.

The author looks at Hydra: five autonomous scout agents find opportunities, Hermes AI makes decisions every five minutes, Jev scores trade setups, and the system learns from its own trades with risk limits hard-coded.

The author sketches a four-line setup where code packs market data into a state, Jev answers three typed questions with probabilities, and code places the order, all in 70 to 500 milliseconds.

The author describes an LLM and Jev trading bot that closed a short and went long without manual intervention when Reliance reversed around 2:30 pm, noting the decisions are AI-made and not a buy/sell recommendation.

The author tagged 44 calls from a support-triage agent run as write or decide, found 37 only picked a branch, and proposed splitting work so Claude writes, Jev decides, and code executes.

When Jev makes an agent's decisions, the question it asked, the answer it chose, and its confidence are all logged, making it easier to see why it read 30 files or skipped tests.

The prompt splits the system into a slow-thinking brain on Opus 5.5 and a reflex layer where Jev returns buy, sell, or hold in about 81ms per candle.

Anthropic's advisor pattern keeps Opus involved only at key checkpoints, while a Jev layer handles over 1,500 mechanical routing branches in 16ms.

Jev is TypeSafe's first System One model, claimed to decide 100x faster and cheaper than a frontier LLM, and this post lists nine tasks you can hand over from Claude.

Sonnet runs the session, Haiku searches, and Opus acts as advisor, with Jev suggested to handle small calls like picking a file or tool or deciding whether to retry.

The plan moves an agent's routine small decisions to Jev and leaves hard problems to Opus 5.5, logging both answers in parallel before switching, with claims of a third of the bill and no rewrite.

The author argues most agent calls are just classification and branching, that Jev returns options with probabilities at roughly 95x lower cost than a frontier model, and outlines a shadow-mode migration.