Simulation infrastructure for agents

Living softwareworlds for agents.

Stateful software populated by simulated users, with years of history, controllable time and exact outcomes.

01 / Stateful software

A running system, not a plausible response.

Deterministic, production-faithful twins with executable APIs, business rules and side effects—not behavior invented by an LLM.

WORLD MODEL / MUTATION PROPAGATION REPLAYING
MUTATION GATE ALLOWED quantity 2 ≤ available 18 COMMIT TX-8421
01STOREFRONT
Order #8419
Order #8420
ENTITYOrder #8421
customer_id
C-204
total
$128
status
PENDINGPAID
02FULFILLMENT
Inventory SKU-72
Inventory SKU-84
ENTITYInventory SKU-91
warehouse_id
WH-3
reserved
4 → 6
available
1816
03BILLING
Payment P-79
Payment P-80
ENTITYPayment P-81
amount
$128
method
VISA · 1842
status
AUTHORIZEDCAPTURED
RELATIONSHIPS
Order #8421contains 2 ×SKU-91paid byPayment P-81
EVENT / RESULT
EVT 01checkout.submittedALLOWED · 3 entities committed

02 / Simulated users

Users observe state, decide and act.

Buyers, admins, employees or operators act from their records, goals and constraints. Their actions create the next work for the agent.

Buyers
Buy, abandon, return or complain
Admins
Approve, defer, escalate or override
Operators
Resolve, reroute, schedule or cancel
POPULATION MODEL / COMMERCE-01 SIMULATING
LIVE ACTIVITY BUYERS · Buyer B-12 placed order ORD-2841

03 / Long horizons

Start with years of history. Run for months.

Start with years of linked activity, then let decisions compound across users, systems and time.

WORLD HISTORY / COMMERCE-01 HISTORY LOADED
SELECTED LAYER Month 34 Mature operations
BUYERS
5,749
ORDERS
125K
TICKETS
1,496
CONNECTED HISTORY
Buyer B-204Order O-8421Ticket T-491Renewal risk

04 / Controllable time

Pause, advance and rewind the world.

Pause, advance, rewind or reproduce the exact conditions around any decision.

WORLD CLOCK / OPERATIONS-01 ADVANCING
WORLD STATE DAY 24 Lower tiers pyramid.lower_tiers_rising

05 / Fork outcomes

Same past. Different agents. Comparable futures.

Fork a world at any checkpoint, run competing agents and compare the business states they create.

CONTROLLED FORK / DAY 43 SAME STARTING STATE
FORK A STATE DIFF / DAY 73 FORK B

06 / Evaluation

Grade the world over time, not only its final state.

Most agent benchmarks stop at task success or final-state checks. Counterworld reveals what changed, when it changed and what compounded next—across users, systems, days and months.

07 / Build any world

Start from a world. Or define yours.

Use a maintained SaaS shape, infer an internal system, or author the complete environment.

01

Catalog worlds

Shopify, Salesforce, HubSpot, Stripe and more.

02

Internal systems

Infer software and actor behavior from APIs, schemas and selected traces.

03

Build from scratch

Define entities, mutations, hooks, event streams and buyer, admin or operator behavior with the SDK.

Give your agent a world that remembers what it did.

Request early access