USE WHEN
When to use this experiment
You are learning how to separate model choice, action validation, and completion checks.
AGENT / SIMULATOR · R03
Simulate an Agent's Next Action
Choose a goal in a fictional settings app and watch Jev select the next allowed action, for up to six steps.
BOUNDED SIMULATORMAX 06 STEPS · NO REAL ACCOUNT
01
MISSION
Configure a fictional inbox
GOALDisable marketing keep security
This target affects only the simulated state shown here.
02
FICTIONAL APP / dashboard
Email controls
NOTIFICATIONS
Marketing emails
Product emails
Security emails
STEP 0 / 6STATE V0● READY
Legal actions: start the sandbox to generate the server-owned action table
HOW THE SIMULATOR STAYS BOUNDED
The server checks each allowed action, tracks the current state, and stops after six steps. The goal is complete only when the simulated settings match it, even if the model says it is done.
LIMITS
What this experiment cannot do
You need to control a real browser, device, or account.
BUILT-IN CHECKS
How this experiment handles edge cases
- No real browser, website, account or network action is executed.
- Pause/reset prevents stale answers from acting.
- An already-satisfied goal returns a local status without a model run or success event.
- Every billed simulation step is counted, including unsuccessful attempts.