Jev Ultrafast: Inside the Speculative Browser Agent
Explain and reason about how Jev Ultrafast drives a browser with one TypeSafe request per decision cycle — from DOM snapshotting through speculative operation/target fan-out to guarded execution — well enough to trace a bug, extend the action space, or evaluate a similar agent design.
5 sections · 16 lessons
Course outline
The core idea: one request, two decisions
- Why browser agents are usually slow
- Speculative fan-out: asking for the answer before you need it
- The action space vocabulary
Seeing the page: snapshot and indexing
- Atomic DOM snapshot (snapshot.js)
- Indexed elements and the numbered element table
- Freshness guards and why retries are dangerous
Deciding: the model layer
- Dynamic operation and target heads (model.py)
- The small text-generation LLM for TYPE_TEXT
- Keeping the model out of the execution layer
Acting: execution and verification
- Executing a target safely (browser.py)
- Waiting without wasting time
- Background tabs, focus emulation, and visible text
- DONE is a claim, not proof
Putting it together and measuring it
- Walking the loop end to end
- What the benchmarks actually show
- Extending the action space without breaking the guarantees