The frontier of agent verification.
Agent Testing Environments
The frontier of agent verification.
Transcripts are stories. Worlds reads the database. Run your agent against a sealed twin of the systems it touches and get exact, replayable proof of what it did.
Stripe, Zendesk and combined workflows. Free internal testing during beta.
● BACKED BY A16Z
ALSO IN THE CATALOG
Stripe
Shopify
Zendesk
Salesforce
ILLUSTRATIVE WORKFLOW
Watch an agent work. Then grade what it did, not what it said.
This synthetic example explains the workflow. It is not a customer result or a performance benchmark.
1. ### Define expected outcomes Start with a saved account and a task. Say which records may change and which must stay protected. 2. ### Run your agent The agent works against an isolated local world using its configured SDK or HTTP client. 3. ### Inspect the outcome Compare the saved expectations with the records that changed. Correct the agent and rerun the same inputs.
WHAT THE AGENT SAID
“The customer’s refund is complete.”
WHAT THE RECORDS SHOW
The refund belongs to a different customer. The expected change is missing.
The dashboard includes clearly labeled scripted samples for billing, native support and combined workflows. Use them to learn the controls, then connect your own agent. Replay guarantees apply to the same world inputs and request sequence; they do not make a model deterministic.
WHAT YOU CAN DO
Six things teams do with a world.
1. Create it A world starts from a seed: the bundled sample account, or a pseudonymized copy of your own. It is ready in milliseconds and identical every time. 2. Mark it A mark records the world's state at that moment. Every diff is measured from a mark, so what your agent did is never mixed up with what was already there. 3. Let the agent act Point your agent's stock SDK at the world and hand it a ticket. Same URLs, same errors, same state machine. No code changes. 4. Read the diff See what changed since the mark: every refund, credit and cancellation, in dollars, and whose money moved. Replay any run byte for byte. 5. Break it Inject rate limits, 500s, lock timeouts, slow responses, an expired key and dropped webhooks. Find out whether your retry logic refunds twice. 6. Move its clock Advance 45 days and let renewals bill, prorations settle, cancellations resolve and webhooks fire in order. Then read the diff again.
HOW THIS IS DIFFERENT
How Worlds fits with the tools you already use.
Keep the tools that help you evaluate your agent. Add controlled system state and explicit outcome checks.
EVAL PLATFORMS
Eval platforms help inspect traces and evaluate outputs. Worlds supplies a local system the agent can act on and a record of the resulting state changes. Use that evidence alongside your existing evaluations.
MOCKS AND SANDBOXES
Vendor test modes and sandboxes have different isolation, reset and simulation features. A world gives each test its own starting state, controllable failures and a logical clock, within the connector's documented scope.
CONVERSATION SIMULATORS
Conversation tests exercise what an agent says to a customer. Worlds lets you check the records its actions change, including changes its reply never mentions.
THE CATALOG
Test the systems your agent uses.
Each one is built to be reset, seeded and broken on purpose, and runs beside your agent, not inside it. When your agent says a task is done, Worlds diffs the environment and tells you what really changed.
● STRIPE AND ZENDESK BETA. REQUEST THE CONNECTOR YOUR TEAM NEEDS NEXT.
PAYMENTS & BILLING
Public beta: the verified implemented billing and payments subset.
COMMERCE
Request Shopify support. Not included in this beta.
GET ACCESS
SUPPORT
Public beta: native support workflows; live API qualification pending.
SALES & CRM
Request Salesforce support. Not included in this beta.
GET ACCESS
world1
/wərld/
I.n.
A working copy of a system your agent acts on. Same endpoints, same errors, same state machine. No real money and no real customers.
II.v.
To run an agent somewhere its mistakes are free.
1 Also see: Sparta, where soldiers were tested before they were trusted.