Coming soon

Describe the job. Get an agent you can check.

Turn your description into an agent. Control what it can do and see what was checked.

.hero-motion_motion__E1kgn[data-hero-motion="pending"], [data-marketing-canvas] > :has(.hero-motion_motion__E1kgn[data-hero-motion="pending"]) ~ *, [data-marketing-canvas]:has(.hero-motion_motion__E1kgn[data-hero-motion="pending"]) [data-hero-body-content] { animation: none !important; opacity: 1 !important; visibility: visible !important; }

Start with the work you want done

Describe the job
Explain the task, the information it needs and the decisions you want to keep.
Set its permissions
Choose what it may read or change, and when it must ask you.
Review the evidence
See repeated task results, attempted misuse and what still needs checking.

A refund agent, with you in control

Start with a request. Shape the tools and permissions around it.

The job you describe

Help with refund requests. Check the order and ask me before returning any money.

The task
Read the request, find the order and prepare a refund.
The tools
Order lookup and a refund tool, with instructions for using them.
The checks
Check ownership, the remaining balance and repeat payment attempts.
The package
Review the tools, instructions, permissions and check results together.

Illustrative refund agent.

The limits you set

Control which records it can read, where data can go and when it must ask you.

Read the right order
Access only the request and order of the authorized customer.
Keep customer data private
Send order details only to destinations you allow.
Ask before paying
Get approval for the exact order and amount. Check again if either changes.
Prevent a repeat payment
Retrying the request must not issue the same refund twice.

Illustrative refund permissions.

See what the checks cover

Proofs cover selected rules in the tool code. Review the rules, assumptions and gaps.

Keep control as it runs

In Axior's runtime, controls sit between the model and its tools. Each call must pass them before it runs.

Catch changed tools
Changed tool definitions or instructions are blocked until they match the approved package.
Keep actions in order
Require steps in the right sequence, such as checking an order before issuing its refund.
Keep private data within bounds
Information derived from private records keeps the same restrictions on where it can be sent.
See why an action stopped
Review tool calls, blocked actions and approvals, with the rule behind each decision.

Know what you can rely on

Review the evidence behind the agent, including where it falls short.

Independently generated tests
Test cases come from your requirements, separately from the agent build. Attacks adapt to its responses to probe for weaknesses.
Reliability beyond a single success
Repeated trials report task success and attack results with uncertainty shown. Checks inspect the actions taken as well as the final answer.
Results tied to a version
Each test record identifies the model, tools, permissions and test cases. Changes require fresh testing before those results apply again.
Gaps you can see
See which requirements are proved, enforced or tested, and which remain uncovered. Disputed test results go to a person for review.

Questions and answers

What verified covers

See which rules and tasks were checked, and what remains uncovered.

Coming soon

Describe the job. Get an agent you can check.

Turn your description into an agent. Control what it can do and see what was checked.

Create an account See how it works

Start with the work you want done

Describe the job
Explain the task, the information it needs and the decisions you want to keep.
Set its permissions
Choose what it may read or change, and when it must ask you.
Review the evidence
See repeated task results, attempted misuse and what still needs checking.

A refund agent, with you in control

Start with a request. Shape the tools and permissions around it.

The job you describe

Help with refund requests. Check the order and ask me before returning any money.

The task
Read the request, find the order and prepare a refund.
The tools
Order lookup and a refund tool, with instructions for using them.
The checks
Check ownership, the remaining balance and repeat payment attempts.
The package
Review the tools, instructions, permissions and check results together.

Illustrative refund agent.

The limits you set

Control which records it can read, where data can go and when it must ask you.

Read the right order
Access only the request and order of the authorized customer.
Keep customer data private
Send order details only to destinations you allow.
Ask before paying
Get approval for the exact order and amount. Check again if either changes.
Prevent a repeat payment
Retrying the request must not issue the same refund twice.

Illustrative refund permissions.

See what the checks cover

Proofs cover selected rules in the tool code. Review the rules, assumptions and gaps.

Keep control as it runs

In Axior's runtime, controls sit between the model and its tools. Each call must pass them before it runs.

Catch changed tools
Changed tool definitions or instructions are blocked until they match the approved package.
Keep actions in order
Require steps in the right sequence, such as checking an order before issuing its refund.
Keep private data within bounds
Information derived from private records keeps the same restrictions on where it can be sent.
See why an action stopped
Review tool calls, blocked actions and approvals, with the rule behind each decision.

Know what you can rely on

Review the evidence behind the agent, including where it falls short.

Independently generated tests
Test cases come from your requirements, separately from the agent build. Attacks adapt to its responses to probe for weaknesses.
Reliability beyond a single success
Repeated trials report task success and attack results with uncertainty shown. Checks inspect the actions taken as well as the final answer.
Results tied to a version
Each test record identifies the model, tools, permissions and test cases. Changes require fresh testing before those results apply again.
Gaps you can see
See which requirements are proved, enforced or tested, and which remain uncovered. Disputed test results go to a person for review.

Questions and answers

What verified covers

See which rules and tasks were checked, and what remains uncovered.

Describe the job. Get an agent you can check.

Start with the work you want done

A refund agent, with you in control

The job you describe

The limits you set

See what the checks cover

Tool rules

Action limits

Task results

Keep control as it runs

Know what you can rely on

Questions and answers

What verified covers

Can I use it today?

Will every answer be right?

What can I review?

What if the agent changes?

What kind of package do I get?

How are permissions chosen?

What is checked before I receive it?

Can an agent be part of an app?

Describe the job. Get an agent you can check.

Start with the work you want done

A refund agent, with you in control

The job you describe

The limits you set

See what the checks cover

Tool rules

Action limits

Task results

Keep control as it runs

Know what you can rely on

Questions and answers

What verified covers

Can I use it today?

Will every answer be right?

What can I review?

What if the agent changes?

What kind of package do I get?

How are permissions chosen?

What is checked before I receive it?

Can an agent be part of an app?