System1 Agents
Run and evaluate fast decision-model agents for browser tasks, games, robotics, and structured choices.
Added to Jevfast
How it uses Jev
The System1 Agents harness exposes Jev as a decision model for one-question choices and fit probes. In its documented blackjack example, it sends player total and dealer upcard state with hit/stand options and a rule to stand on 17; the selected choice is returned with per-option probabilities, confidence, and latency.
What Jev decides
Illustrative blackjack choiceExample answers · not a recorded Jev response · Source ↗
Question 1 · pick
YOUR APP
INSTRUCTION
Choose between hit and stand under the supplied rule to stand on 17.
STATE
Player total 18 and dealer upcard 9, from the documented CLI example. The stated rule is to stand on 17.
JEV · CHOICE
- hit
- stand
App workflow
- Prepare a documented case
The CLI accepts JSON state, repeated key-description options, and a rules string.
- Ask the decision model
The probe sends one ChoiceQuestion under the key `pick`; the Jev adapter serializes the state and question to the configured decisions endpoint.
- Measure task fit
Across hand-written cases, the probe compares the selected option with accepted keys and reports accuracy, median latency, and a fit verdict.
Why it is interesting
It provides a small probe for whether a task fits a fast choice model, and reports accuracy and latency across hand-written cases before calling that task a fit.