Jev in Practice: Browser Use and Agent Workflows

Anthropic's Boris Cherny lays out a Claude Code setup where Sonnet 5.5 runs the main loop, Haiku 5.5 handles parallel read-only scouting, and Opus 5.5 advises only at plan lock, repeated errors, and pre-merge; Jev is framed as the same split one layer down for millisecond forks.

An open-source workflow streams the board from Ray-Bans to a Mac, where Jev cheaply filters unchanged frames before Opus 5.5 reads the position and Stockfish suggests a move spoken into the call.

The author uses Jev to scan wallets around a token and write a one-line risk reason per wallet, while Grok Bot blocks, sells, or skips; the demo shows at-risk exposure dropping from $4,820 to $0 in 33 seconds.

A post relays Anthropic's Boris Cherny on splitting Claude Code across Sonnet 5.5, Haiku 5.5 and Opus 5.5, and argues that low-stakes forks like which file or tool to use can go to Jev in milliseconds.

A list of 32 Grok Bot skills adapted from GitHub playbooks, with jev-passage-filter ranking retrieved passages and jev-typed-decision handling small routing and classification calls.

A developer's end-of-year schedule includes a Nov 5 San Francisco workshop with Mastra on using judge models such as Jev to build agents that handle sensitive data.

Lightpanda's agent uses Jev to pick the username, password and submit fields in 0.53s with 2 calls, versus Claude Haiku 5.5 at 7.83s and 8 tool calls.

A shared Jev prompt asks an agent three yes/no questions per page — unclear problem, out of date, near duplicate — to decide fix, merge or hold; a first pass flagged 2,192 of 3,208 pages until a scraper bug was fixed, dropping it to 666.

Configures Sonnet as the main model, Haiku 5.5 as read-only scout and Opus 5.5 as advisor per Anthropic docs, suggesting Jev handle small choices like file, tool and retry selection.

The author says Jev lets you teach Auto Router preferences like preferring GPT 6.1 Sol for copywriting over Opus 5.5, in a few clicks.

The team says it built the first Decision API grounded in web data to feed Jev's System One models, at 10x lower cost than a standalone model.

A short reply advises combining typesafeai's Jev with Claude in a hybrid setup, claiming noticeably better results, though it offers no concrete workflow or numbers.