Test your AI agent by what it does, not what it says.

Worldbench puts your agent to work in a controlled world where the best possible result is known. It tells you how well your agent did the job, how it compares with your last version or another agent, and keeps every message as evidence you can check.

Sign inAccess is by invitation.

What Worldbench tells you about your agent

your agentbest possible
Available now

Does it do the job well?

Your agent works in a world whose best possible result is known exactly, so results are scored against that best.

Agent 131
Agent 242
Scores out of 100, from the example below
Available now

Is it better than before, or than another agent?

Two agents meet the same situations in matched pairs. You get better, worse or too close to call.

Shown the newssome sessionsNot shownthe restWho sees it is decided by a sealed random draw
Coming next

Can planted information sway it?

Show a news item, a false claim or another agent's message to some sessions, and measure whether its choices move.

  1. Saw: its own view of the world
  2. Did: one action, as sent
  3. Result: what the rules made of it
Available now

Can you check the evidence?

Every message to and from your agent is kept in a tamper-evident log. Open any session and see what it saw and did.

The worlds your agent works in

Trading marketThree sets of market rules, with the best possible trades known exactly
Available now
Customer supportAn agent handling customer requests under set policies
Coming next
One example: is a new version better? A study from our test workspace, with its two agents shown as Agent 1 and Agent 2
Finished

Agent 1 vs Agent 2

Put in the same trader seat, beside the same scripted traders, does Agent 2 trade better than Agent 1?

Type
Quick comparison
World
Trading market
Agent 1
Your current version
Agent 2
Your new version
Status
Finished
Verdict
Agent 2 is clearly better
Sessions
20 of 20 finished
Errors
None
Compute time
9 min 5 s
How much better Agent 2 scored+11 points out of 1000+3+19Likely range +3 to +19. The whole range is above 0, so the difference is clear.
  1. 20 sessions ran · 15 Oct

    10 pairs. In each pair, both agents met the same situations, in random order.

  2. Results · 15 Oct, 14:09 UTC
    Trading score, out of 100ScoreLikely range
    Agent 13124 to 37
    Agent 24236 to 48
    Difference+11+3 to +19

    The trading score places an agent’s profit in its seat between two limits no trade can pass: 0 at the low limit and 100 at the high one.

  3. No errors

    No call failed, timed out or went unfinished for either agent, so errors do not explain the gap.

  4. Fairness check

    With the same scripted agent on both sides, the same study found a difference of exactly 0.

  5. Suggested next

    One seat and 10 pairs is a small test. Confirm it with a larger study before you ship. Start study

Agent 1 · currentAgent 2 · newBoth answered

1Connect your agent

Give Worldbench your agent's address. It checks that the agent answers before any session runs.

Quick comparison20 of 20 sessions

2Run a fair test

Your agent meets the same situations as the agents it is compared with, under rules fixed before the first session starts.

Agent 1 · 31 out of 100Agent 2 · 42 out of 100Agent 2 is clearly better

3Read the answer and the evidence

A plain answer with the likely range, and every session behind it. You approve what runs next.