Jev — The Fast Decision Engine
Jev is the decision engine that drives browser actions fast: one small-model decision per action, no full-model round-trip per click, no screenshots in the decision loop — just the accessibility tree.
The problem Jev solves
The naive browser loop works like this: screenshot the page → send it to your big main model → wait for it to decide → click → repeat. In real sessions this was painfully slow and context-heavy — tool-result tokens ballooned past 290K in a single session, and every click cost a full frontier-model call.
Jev flips the loop into jev-ultrafast style:
- Snapshot — the browser produces a compact accessibility tree (buttons, links, inputs as
@e1,@e2, …). - One Choice call — a small, fast model sees the element table + your goal and returns exactly one action (~1–2 seconds).
- Execute — AgentX performs the CDP action (click, type, scroll…), waits briefly for the page to settle, and repeats.
Your expensive main model never touches micro-decisions. It only sees the summary when the task finishes.
How decisions stay safe
- Structured element table — the engine picks from indexed elements (
@e1,@e2), not raw coordinates, so clicks land on real controls. - Action-specific targets — each action type (click, type, select…) gets the right element kind.
- Quick post-action waits — ≤200ms settle after typing, so the next snapshot reflects reality.
- Loop protection — an action that changes nothing is blacklisted and never retried; the same action 3× in a row stops the task instead of spinning forever.
- Confidence gate — below a minimum confidence threshold the engine stops as "uncertain" rather than clicking blindly.
- Fail-open — if Jev isn't configured or errors mid-task, the main model takes over. Jev never breaks a user-visible flow.
Setup
- Get a Jev-compatible API key from your Jev provider.
- In AgentX, open Settings → Jev (its own settings page — separate from model providers, like web search).
- Paste the base URL and model name yourself; no model picker needed.
- Paste your API key — multiple keys, one per line, are supported and the agent rotates across them to spread load and avoid rate limiting.
- Make sure Jev is the selected decision engine.
When the main model steps in
The main model is used only as a fallback — when Jev is not configured, the key is missing, or a decision call fails. If Jev is healthy, the model never touches browser micro-decisions. You can watch which engine is active in the browser watch panel.
Jev vs the alternatives
| Jev | Drex | Main model | |
|---|---|---|---|
| Speed per action | ~1–2s | ~1–2s | Slow (full round-trip) |
| Cost per action | Tiny | Tiny | Full model call |
| Page context | Compact element table | ~4× larger table | Whole conversation context |
| Setup | Bring your own key | Invite-based | None |
See the full comparison with recommendations: Browser Automation Options.
What's next
- Drex — the higher-context alternative engine
- Browser Automation Options — backends vs engines, what to pick