A supplied run, before-and-after state, preconditions, postconditions, expected changes, observed effects, and policies.
AI agent outcome verification API
Verify whether an AI agent actually achieved its outcome
Supply the claimed result with before-and-after world state, assertions, side effects, and policy. Outcome Assurance returns what held, what failed, and why.
Free: 500 assured runs/month. Developer: $249/month with 10,000 included. Compare allowances and pricing.
Exact use case
An agent or automation reports success, but downstream systems need independent proof that the intended state changed without forbidden side effects.
Assertion results, undeclared or forbidden side effects, regressions, a verdict, evidence, and a receipt.
Evaluates explicit outcome assertions and side-effect policies against supplied state and evidence.
A successful authenticated POST /v1/runs that returns a completed assurance verdict.
From the problem to a useful result
Choose the part of the review you need to solve
Verify the outcome against supplied state
AI agent outcome evaluation checks whether the supplied result satisfies the task’s stated conditions. Bring before-and-after state so the review can distinguish a claimed success from the state changes actually recorded.
The refund example includes the order, customer tier and ledger state. The API evaluates those supplied records; your application is responsible for capturing them.
Make success conditions explicit
Define agent preconditions and postconditions as assertions. The example requires the order to start as paid and finish as refunded with a refund identifier present.
Inspect the equals and exists assertions in the demo. Each assertion produces an explained result, making the success definition reviewable without a subjective model grader.
Detect forbidden changes and undeclared effects
Agent side-effect testing also asks what else changed. Declare expected effects and supply observed effects, then use explicit policies for changes the task must not make.
In the synthetic example, the refund conditions hold, but the customer tier changes under a prohibition and an email is undeclared. Inspect the outcome verdict together with the policy and effect findings.
Reuse the checks across runs and releases
Apply the same outcome assertions when reviewing later runs or evaluating an agent release. An optional baseline supports regression comparison within the supplied run data.
Create and verify a free account to call /v1/runs from your own harness. Your system collects state and decides what to do with the verdict; outcome checking does not itself block a run or roll back an action.
OpenAPI-grounded keyless demo
Assure the run
Edit the supplied example, send it to /v1/demo/assure, and inspect the service’s unedited JSON response. The demo stores nothing and consumes no account quota.
Response
Run the demo to see the API response.
A successful response demonstrates the bounded classification or calculation shown here. It is not evidence of a deployed connector, autonomous external action, customer outcome, or commercial performance.
Pricing and included usage
Choose a plan for recurring outcome checks
One supplied run evaluated in one call is one assured run. A batch of 50 runs uses 50 units; an attached baseline comparison is not billed as a second run.
Free
$0/month
500 assured runs/month
No card required. The keyless demo consumes no account quota.Developer
$249/month
10,000 assured runs/month included
$0.05 per additional assured run. Overage charges capped at $500 per billing period, in addition to the base price.Growth
$999/month
60,000 assured runs/month included
$0.03 per additional assured run. Overage charges capped at $2,000 per billing period, in addition to the base price.Scale
$2,499/month
250,000 assured runs/month included
$0.02 per additional assured run. Overage charges capped at $5,000 per billing period, in addition to the base price.Prices in USD, billed monthly. Choose a plan for the volume you expect to run. See full pricing and Enterprise options.
Capability boundary
What this page does not promise
It is not a model benchmark, subjective LLM evaluator, or generic HTTP response assertion tool, and it does not observe systems you do not describe.
- The demo qualifies the transformation only; account verification, key creation, and a successful authenticated primary-value request are required for activation.
- Inputs and outputs must be reviewed in the context of your own data rights, policies, and operational controls.
- No ranking, accuracy, savings, recovery, prevention, customer, or revenue result is claimed.
From evaluation to first value
A short, explicit path
- Run the keyless demo.
Confirm that the input and output shape match the job you need done.
- Create and verify a sandbox account.
The email claim flow creates a product-scoped API key. No card is required for the free tier.
- Call the authenticated primary-value operation.
A successful authenticated POST /v1/runs that returns a completed assurance verdict.