Start from your actual rules.
Set the refund window, automatic limit, currency, tool names and policy version. Inspect the assumptions before approving a case.
EVAL WRITING / SUPPORT AGENTS
Your agent passed
the demo.
Now test the edge.
Turn your support rules into concrete eval cases. Catch the wrong refund, the repeated action and the confident answer with no evidence.
Free sample. No account. Your policy stays in your browser.
One cent over.
A different decision.
A refund limit is only useful if the agent respects its boundary.
- Automatic limit
- $100.00
- Requested refund
- $100.01
- Expected outcome
- Escalate to a human
- Permitted money movement
- $0.00
assert refund.success_count == 0
assert state.refund_cents == 0
assert state.refund_cents == 0
Illustrative policy and synthetic fixture. No real agent result is shown.
32 editable patternsObserved-state checksJSONL + Markdown exportsOffline writer + Node grader
01 / THE DELIVERABLEA useful eval starts with
A useful eval starts with
a precise expectation.
Write the dataset and the checks. Keep the agent harness and model you already use.
Policy → fixtures
Make failure observable.
Check exact amounts, ownership, successful side effects and attempted actions. Keep qualitative review separate.
Take the suite with you.
Export editable JSONL and a readable specification. Grade observations from your sandbox adapter with the local grader.
02 / WHERE AGENTS GET STUCKBeyond
Beyond
the happy path.
32 original synthetic patterns across ten categories. A starting library to adapt to your product, with its exact scope visible in every case.
- ↳Refund policy & exact boundaries
- ↳Identity & resource ownership
- ↳Duplicate requests & idempotency
- ↳Failed calls & unknown outcomes
- ↳Cancellation & withdrawn consent
- ↳Missing evidence & invalid amounts
- ↳Private data & untrusted content
03 / OWN THE WORKBENCHOne purchase.
One purchase.
A suite you can keep.
Both editions include the complete 1.0 workbench. Choose the license for how you’ll use it. The complete workbench is available as an instant digital download.
For your own agents.
$199 USD / once
- 32 editable scenario patterns
- Offline writer and observation grader
- JSONL, Markdown and source files
- Incident-to-eval writing workflow
- Internal use by one organization
For the agents you deliver.
$599 USD / once
- Everything in Team
- Use across your client projects
- Deliver adapted suites to clients
- Include the grader in client handoffs
- One agency; unlimited client projects
No subscription. Model usage, custom consulting and managed monitoring are not included. This purchase covers edition 1.0; no future update schedule is promised. 14-day refund requests. License & purchase terms.
Does this use AI to write cases?
The workbench uses explicit, editable patterns and policy parameters. It does not send your policy to a model or infer arbitrary business rules. An optional assistant prompt helps draft additional cases from redacted incidents for review.
Does it run my agent?
It writes cases and grades normalized observations. You connect your agent to sandbox tools and capture results using the adapter contract. The Node grader needs Node.js 22 or later; the browser writer works offline.
What does a pass mean?
A mechanical pass means the supplied observations satisfy the listed checks. Review policy correctness, response quality, privacy and activity outside those checks separately. It is not a certification or complete production-readiness assessment.
Can I use my existing eval platform?
You own editable neutral JSONL and source. Map the documented schema to your harness or platform. A one-click third-party integration is not included.
Does my customer data leave my computer?
Writing and grading happen in your browser or local Node runtime. There are no model calls or trace uploads. Purchase and account details are processed separately by the hosting service and Stripe.