counterworld Request early access

Solutions

Test agents against consequences.

Train and evaluate agents—or specialize a smaller model—inside stateful software where users keep acting, time advances and decisions compound.

Request early access

01 / Agent training

Train on state transitions and delayed rewards.

Let an agent practice inside resettable worlds where valid actions change software state and simulated buyers, customers, admins, employees or operators respond.

Start
A checkpoint with realistic history
Run
Agent, customers, admins and operators act on shared state
Reward
Outcome quality, rule adherence and cost

02 / Small-model specialization

Train a smaller model to perform like a larger one—on your work.

Build environments from your organization’s interfaces, policies, user behavior, history and edge cases. Generate verified frontier-model rollouts, then fine-tune a smaller model to meet or exceed that quality on a bounded workflow with lower cost and latency.

Environment
Your systems, users, policies and failure modes
Training signal
Frontier-model rollouts scored against resulting world state
Result
A smaller model specialized for a defined operational workflow

03 / Long-horizon evaluation

Run work that takes weeks, not turns.

Advance the world for 30 or 90 days while customers, support agents, billing admins, warehouse operators and schedules continue producing new conditions.

Task
Run support operations for 90 days
Pressure
Tickets, renewals, refunds and churn
Score
SLA, backlog, retention and rule breaches

04 / Regression testing

Turn a failure into a deterministic test.

Capture the world before a failure, replay the same future workload and compare the resulting state after every agent change.

Capture
Checkpoint before the failed decision
Replay
Same customers, admins, operators, schedules and incoming work
Assert
Failure gone and invariants still hold

05 / Agent version comparison

Compare agent versions from identical starting state.

Fork one checkpoint, then run agent variants against the same buyers, customers, admins, employees or operators. Change the prompt, skills, tools, memory or policy and measure which version creates the healthier world.

Variants
Prompts, skills, tools, memory and policies
Control
Same history, actor roles and future events
Compare
Outcome, violations, stability and cost

Shared scoring

Score the world after the agent acts.

Every workflow evaluates state and outcomes, not only the agent’s final response.

01

Task outcome

Was the assigned work completed?

02

Business outcome

What changed for users and the operation?

03

Rule violations

Which permissions or invariants broke?

04

Recovery

Did the agent detect and repair failure?

05

Stability

Did performance hold as conditions changed?

06

Cost

How much time, compute and intervention?

06 / Production replicas

Recreate production systems and their users without cloning production.

Use internal API contracts, schemas and selected traces to recreate both the software and the people using it: buyers, customers, admins, employees, approvers and operators. The software maintains shared state, hooks, event streams and side effects while these users make decisions and encounter failures, without any production database copy.

Learn from
Internal APIs, schemas and sanitized production traces
Recreate
Entities, mutations, hooks, event streams and user decisions
Inject
Synthetic roles, traffic shifts, policies and system failures

Run a scenario

Start with one workflow.

Choose a world, define the population and horizon, then connect the agent and evaluators.

01 Choose or create a world 02 Define the task and scoring 03 Run, inspect and replay
Request early access