Adnan Ahmad: A fast browser agent with Jev and Stagehand

Traditional browser agents send thousands of tokens of HTML to a large model for every DOM interaction, costing 3-8 seconds and $0.05-$0.20 per step. System One splits each step: sub-150ms typed reflex decisions with Jev, and deep reasoning with Gemini or Claude only when needed.

target (choice): which candidate action should be executed next; is_complete (noul): has the user's objective been fully achieved; is_destructive (noul): would this action trigger a payment, deletion or credential change; and page_category (choice): search, form, content or anti-bot.

When Jev's confidence drops below 0.55, or when the task requires complex creative text synthesis such as filling forms with generated content or composing messages. The harness supports Gemini Flash and Claude Sonnet as fallback providers.

is_complete above 0.82 terminates with success; is_destructive above 0.75 halts with a safety warning before the action is dispatched; confidence at or above 0.55 takes the fast path; and a stagnant page fingerprint for three steps or consecutive anti-bot categories ends the run.

Because comparing a sub-second reflex against an entire multi-step workflow is misleading. It reports three metrics separately: Jev reflex at roughly 70-150ms, Stagehand observe() at 1.5-3.5s, and total step at 2.0-4.5s including DOM settlement and CDP screencast.

No. Jev replaces the decision LLM call, not the observe() call. Every step still pays a Stagehand observe() round-trip to discover actionable elements before Jev is consulted, and the fast path executes a direct Playwright selector click with no LLM generation.

A Next.js 16 and React 19 UI with live viewport screencasting over WebSocket from Chromium's CDP, a step timeline with candidate rankings and confidence meters, a step inspector for Jev answers and probability distributions, live telemetry, and keyboard shortcuts. Its bridge binds to 127.0.0.1 and validates the Origin header.