JEVANY / AGENT HARNESS

The same task.
Fewer LLM calls.

Watch the LLM and Jev take turns. Compare recorded calls, token use, completion time, and the task outcome.

LLM Jev Task result

THE OVERVIEW

Three tasks. One replay.

Download overview video
WebShop → FrozenLake → SQLite. Three seconds per case, using the same replays shown below.

THE INDIVIDUAL DEMOS

See where Jev takes over.

01 / WEBSHOP

Bound the choices.

The LLM provides two menus. Jev selects the requested color and size; the LLM completes the purchase.

D4 supplement · seed 3107 · one recorded pair

Download video

02 / FROZENLAKE

Keep the plan.

One LLM objective stays in place. Jev chooses a direction from each new state, making four local moves to reach the goal.

Seed 3000 · one recorded pair

Download video

03 / SQLITE RECOVERY

Choose the evidence.

The LLM proposes three inspections. Jev selects raw-page inspection; the LLM recovers ten rows, and an independent verifier checks the output.

Terminal-Bench 2 · sqlite-db-truncate · one recorded pair

Download video

READ THE RESULTS IN CONTEXT

One replay shows one recorded pair.

Calls, tokens, total time, and rewards come from the recorded runs. Savings use the LLM-only baseline. Playback preserves each pair’s measured completion-time ratio; individual step durations are illustrated.

Broader results vary by task. The report includes the original D2 matrix, the separate D4 WebShop supplement, fallbacks, and uncertainty.