Counterworld platform

FORKa living software world. RUNan agent. COMPAREoutcomes.

Counterworld combines stateful software, grounded simulated users and controllable time in one runtime. Run current and candidate agents under identical conditions, then compare the business state each one creates.

Request early access View the SDK

RUN / compare-agents Illustrative run

fork(cp_day_00)

agent@v17complete

horizon: 30d
events: identical
output: world_a

agent@v18complete

horizon: 30d
events: identical
output: world_b

comparisoncandidate leads

invariants18 / 18 pass

artifactworld.diff.json

01 / How it works

One controlled run from starting state to business outcome.

The agent operates through normal interfaces. Counterworld controls the world, its clock and the experiment around it.

PIPELINE / release-evalIllustrative · 6 / 6 complete

01 / CREATEdone
worlds.create(definition)→ world_8f2
02 / CONNECTdone
agent.connect(scoped_endpoint)→ credential_agent
03 / RUNdone
world.run(agent@v17)→ trajectory_a
04 / ADVANCEdone
clock.advance({ days: 30 })→ checkpoint_day_30
05 / FORKdone
world.fork(checkpoint_day_00)→ world_candidate
06 / EVALUATEdone
compare(world_a, world_b)→ report.json

02 / Inside a world

Complete interfaces. Exhaustive behavior.

For every workflow in scope, the API surface, business rules and mutations are fully executable—not generated call by call.

01 / API surface

Production-shaped interfaces

Endpoints, schemas, authentication, pagination and error contracts.

02 / Business semantics

Every valid consequence

Permissions, lifecycle rules, invariants and cross-system side effects.

03 / Failure behavior

The unhappy paths too

Idempotency, retries, timeouts, rate limits, partial failures and 4xx/500 responses.

WORLD INSPECTOR / commerce-opsIllustrative · clock running

State / entities

Entity	Records
customers	5,749
orders	125,204
payments	128,901
fulfilments	119,442
tickets	1,496

Event stream / latest

09:41:02
checkout.submitted
Order #8421 enters mutation gate
09:41:03
payment.captured
Payment P-81 commits $128
09:41:03
inventory.reserved
SKU-91 available 18 → 16
09:42:10
buyer.message
Customer C-204 requests an update

Population / active

Buyer / C-204Maya ChenWaiting on split fulfilment
Admin / A-17Dev Patel3 approvals due today
Operator / O-04Leah GrantWarehouse shift ends 18:00

History / 3 yearsDAY 11 · 09:42:10

03 / Controlled comparison

Change the agent, not the conditions.

Forking creates two futures from the same checkpoint. Both agents face the same history, users, permissions, schedules, failures and external events.

EXPERIMENT / agent-releaseIllustrative comparison

checkpointcp_day_00

seedsim_042

horizon30 days

workloadidentical

Baseline

Current agent

complete

model: large-v4
prompt: support-17
memory: retrieval-v2
policy: refunds-4

SLA91%

Refund errors2

LedgerPASS

Candidate

Agent under review

complete

model: small-v7
prompt: support-18
memory: retrieval-v3
policy: refunds-4

SLA96%

Refund errors0

LedgerPASS

04 / Evaluation output

Grade the world, not only the final response.

Every run produces evidence at the task, business, state and trajectory levels so teams can see whether an agent improved—and what it broke along the way.

REPORT / world_a..world_bIllustrative report

Business outcomes

Metric	Current	Candidate	Delta
Resolution SLA	91%	96%	+5 pt
Open backlog	184	142	−42
Duplicate refunds	2	0	−2
Human interventions	38	31	−7
Compute cost	1.00×	0.71×	−29%

Invariant checks

refund_total ≤ capturedPASS
inventory.available ≥ 0PASS
ledger.debits = creditsPASS
notification.sent_oncePASS
role.can_mutate(resource)PASS

world.diff / order_8421
- refund.status: "pending_review"
+ refund.status: "approved_partial"
+ refund.amount: 42.67
+ audit.policy: "refunds-4"
  ledger.balanced: true

05 / Integration

Control the experiment without changing the agent.

Use the Counterworld SDK to create and operate worlds. Give the agent only a production-shaped endpoint and scoped credentials for its instance.

Agent plane

Production-shaped access

API, SDK, MCP and tool interfaces with role-specific permissions.

Control plane

Privileged experiment access

World creation, clock control, checkpoints, forks, full-state inspection and evaluators.

Open the SDK quickstart

experiment.ts

const world = await counterworld.worlds.create({
  systems: [
    catalog.shopify(),
    catalog.stripe(),
  ],
  history: { years: 3 },
  population: [buyers(), storeAdmins()],
  evaluators: [
    noDuplicateRefunds(),
    balancedLedger(),
  ],
})

const start = await world.checkpoint()
const current = await world.run(currentAgent, {
  days: 30,
})

const candidateWorld = await world.fork(start)
const candidate = await candidateWorld.run(
  candidateAgent,
  { days: 30 },
)

const report = await compare(current, candidate)

06 / Trust and boundaries

Controlled rehearsal, with explicit limits.

Counterworld provides isolated, inspectable worlds for experimentation. It does not require production access by default or claim that simulated behavior predicts real human conversion.

Start with one workflow

Define the world required to judge the outcome.

Choose a consequential workflow, identify the state and actors that affect it, then run current and candidate agents under the same conditions.

Request early access Explore world creation

World isolation: Separate state, credentials, clock and lifecycle for every run.
Data source: Synthetic histories and populations by default; production data is not required.
Replay: Reproducible when starting state, seeds and future workload are fixed.
Evaluation access: Privileged state inspection remains outside the agent boundary.
Interface fidelity: Coverage is explicit for every modeled API, SDK, MCP or tool surface.
UI fidelity: API fidelity does not automatically imply a faithful browser interface.
Interpretation: Behavioral parameters define controlled test conditions, not human forecasts.
Production role: Simulation supports release decisions; it does not replace production monitoring.