Agentic Watch
Independent testing of your agent software on every release, signed by a named person, written so your team can use it to answer security questionnaires.
If you build software that acts on its own (an agent, an agent runtime, a gateway, an MCP server, a workflow engine or a developer tool), your customers' security teams now ask how it behaves. What can it reach? Does it stop when told no? Where does its data go? Can you show it still holds after your last release? Agentic Watch tests those things on each release you ship and gives you a signed record you can attach.
How it works
- Setup. We read your documentation and security claims, agree the tests with you, and build a test environment for your product. Setup ends with your first signed record.
- Every release. When you ship, you tell us the version. We install that exact build in a fresh, isolated environment, run the agreed tests, and send you the signed record.
- Every quarter. A short summary of every version tested, what passed, what failed and what was fixed, so you can show the tests ran over time and not just once.
What you get for each release
- A signed record for that exact version: version, commit, date and a fingerprint of the file so anyone can check it has not changed.
- Each test marked pass, fail, does not apply, or could not be tested in our environment, with the evidence behind it.
- Anything that changed since your last release, flagged clearly.
- For every failure: the steps to reproduce it. Once you have fixed it, we re-check it.
- A one-page answer sheet with wording you can paste into a security questionnaire, each answer pointing to the record.
- A machine-readable copy you can upload to your compliance tool.
The tests
Watch 6
Watch 6 tests the six things that most often stop a deal:
- Refusals and approvals hold when the agent is asked directly to get round them.
- They still hold when the instruction is hidden in a file, a tool result or a web page.
- It stays inside the files, processes and tools it is allowed to use.
- It only contacts the services its documentation names.
- Secrets in its environment never leave it or appear in its logs.
- What it sends back to its maker matches what its documentation says it collects.
Watch 12
Watch 12 adds six more:
- It stops when told to, and we measure how quickly.
- Its credentials and permissions cannot do more than documented.
- Access that has been revoked is refused on the next action.
- Every action it takes appears in its own audit log, under the right identity.
- and 12. Two checks written from your own published security claims.
Watch 6 does not cover stopping, credentials, revocation, audit logs or your own claims. If your customers ask about those, choose Watch 12.
Prices
| Plan detail | Watch 6 | Watch 12 |
|---|---|---|
| Tests per release | 6 | 12 |
| Releases tested | 6 a quarter, up to 3 in any month | 6 a quarter, up to 3 in any month |
| Monthly | £1,500 | £2,500 |
| Annual (12-month commitment, 10% off) | £16,200 | £27,000 |
| Setup, one-off, includes your first record | £2,000 | £3,500 |
| Record delivered within | 10 working days of your release notice | 7 working days of your release notice |
- Minimum term six months. Billed quarterly in advance.
- These prices are for products that install from one image or installer. If yours is several services working together, we quote setup and any monthly difference after a one-hour look.
- Need more than 12 tests or more releases? We quote it.
- Unused releases expire at the end of the quarter.
- All prices are the total you pay. The fee does not depend on the result.
Where your code runs
Each run uses a fresh virtual machine in London, built from our hardened image and deleted when the run ends. Your code and builds exist only there, for that run. It has no network access except to install dependencies and to reach the model through our proxy, so your real keys never sit inside the test. We hold no credentials to your live systems and never test them. Any AI we use is from a provider whose terms forbid training on your data.
What it is not
Not a certification, an audit opinion or a penetration test. Not a test of your hosted service, cloud setup or live customers. Not an endorsement. You may say "independently tested by Agentic Thinking Ltd at version X on date Y, N of M checks passed" with a link to the record. You may not say certified, secure, audited or compliant because of it.
We hold professional indemnity and cyber insurance and are registered with the ICO.
Questions
- Can we choose which releases are tested?
- Yes. You tell us which ones, up to your allowance.
- What if something fails?
- You get the reproduction steps the day we sign the record, and one re-check of all your fixes within 30 days, which does not use up a release.
- Can we switch between Watch 6 and Watch 12?
- Yes, at the start of a quarter. Moving up costs the difference in setup.
- Is there a public page?
- Only if you want one. It shows each version tested, the date, the score and each check by name, never the details of a failure.