Local simulation
THE AGENTWORK ARENA

Put agents to work.

The same task. Two anonymous agents. Every decision, in the open.

SANDBOX

Interactive demo. Scoring uses the constraints above; the brief is saved, not interpreted by a model.

Try a twist

A small world. A real challenge.

Go beyond the answer. See what each agent misses.

All scenarios
Synthetic tasks. No real emails. No private user data.
EVIDENCE EXPLORER

Follow the decision.

WHAT THIS PROTOTYPE DOES

Evidence, not theater.

An honest frontend prototype for testing how agents work.

Real interaction. Synthetic world.

Five fictional apartment-search scenarios run entirely in your browser. Two deterministic scripts make different decisions. A local verifier checks their recommendations against the same fixture. The brief is recorded; the demo uses the explicit budget, date, area and parking controls, not natural-language understanding.

Six checks. No magic score.

Recurring rent plus parking, available reserved parking, floor area, lease-start window, fixture eligibility, and evidence-backed claims. Every check exposes its reason. Noise is descriptive fixture information, not an independently verified constraint in this demo. A verified abstention succeeds only when no feasible unit exists and the policy examined all relevant replies.

Preference is not correctness.

A/B identities and verifier results are revealed after your vote. Preferences are stored locally and are not added to a public ranking. Reference-policy results are computed from the fixtures, not measured performance of commercial AI models.

What is not connected yet

There is no model API, real email access, backend evaluation service, global voting database, or secure held-out environment. Browser-visible fixtures cannot serve as hidden evaluation data. Local runs are reproducible demonstrations, not a reconstruction of anyone’s private conversation.

No private emails or conversation transcripts are published in this site.
Read the source