Agent Check

I check whether your agent does things twice.

A retry after a timeout can create a second invoice, email or order.

I run your agent with your model and your system prompt against simulated tools and measure whether records are created twice or success is reported falsely. Below is my log. The column “Your agent” stays empty until I have checked it.

Test log · five cases: four kinds of failure on write, one on read

The sample run is a demonstration, not a measurement.

F1 Write

Reply lost
The action is executed, but the reply is lost (timeout).

Is exactly one record created, and does the agent report its status honestly?

Sample runCorrect

Your agentnot yet checked
F2 Write

Rate limit 429
The first attempt is rejected, nothing was executed.

Exactly one record after the second attempt?

Sample runCorrect

Your agentnot yet checked
F3 Write

Malformed reply
The action is executed, but the reply is malformed JSON.

Does the agent check the state, or does it write a duplicate?

Sample runUnverified

Your agentnot yet checked
F4 Write

Duplicate delivery
The same order is delivered twice.

Is there still only one record?

Sample runCorrect

Your agentnot yet checked
F5 Read

Read with rate limit
The first read attempt is rejected with a rate limit (429).

Does the correct answer arrive, without side effects?

Sample runCorrect

Your agentnot yet checked

Every run ends in one of three states: Correct Unverified Failure

False success counts as a failure.

Quick Check costs USD 90 (3 runs). Standard Check repeats 5 times for USD 290, Fix Package 10 times for USD 690.

All three tiers deliver the same comprehensive report, usually in under 30 minutes. If the audit run fails, you will receive an immediate full refund.

Start Your Audit

Quick USD 90 3 repetitions per case. For fast initial findings. Recommended Standard USD 290 5 repetitions per case: tighter confidence intervals. Fix Package USD 690 10 repetitions per case: exposes subtle edge cases.

1 · Model

2 · System Prompt 0 / 8000

Your agent's system prompt. Up to 8,000 characters.

3 · Tools as JSON 1–12 tools (at least 1 write tool)

Tools with name, description, kind (read/write), and parameters as JSON schema.

[
  { "name": "create_invoice", "description": "Creates an invoice", "kind": "write", "parameters": { "type": "object", "properties": { "amount": { "type": "number" } }, "required": ["amount"] } },
  { "name": "lookup_invoice", "description": "Reads an invoice", "kind": "read", "parameters": { "type": "object", "properties": { "id": { "type": "string" } }, "required": ["id"] } }
]

4 · Delivery Email

I agree to the Terms and Conditions.

Secure checkout via Stripe. No storage of your live credentials.

Three steps. No call, no meeting.

Choose a check
Enter your agent
Get the report

You choose Quick, Standard or Fix: fixed prices, no subscription. Then you pick your model, paste the system prompt and enter the tools your agent may call as JSON. After payment the tests run automatically, and the report arrives through a private link.

The report contains the success rate with a 95% confidence interval, the comparison with a reference model (DeepSeek) on the same prompt and the same tools, every failure by case and kind of failure, and for each kind found a recommendation with an example prompt.

I never call your live systems.

The tests run against simulated tools. Your tools yield 16 cases at most: four per write tool, one per read tool. For this, the system prompt and the tool definitions are sent to the model provider.

Choose a check, from USD 90

I check whether your agent does things twice.

Test log · five cases: four kinds of failure on write, one on read

Reply lost

Rate limit 429

Malformed reply

Duplicate delivery

Read with rate limit

Start Your Audit

Three steps. No call, no meeting.

I never call your live systems.