Harness Watch
Private pre-release runs of the AgenticBench security suite, for companies that ship coding agents.
Harness Watch runs the AgenticBench security suite on your coding agent before you ship it. It is the same set of behaviour checks behind our public results: whether the agent stays inside its permissions, stops when told no, keeps credentials and data where they belong, and resists hostile content in repositories. We run it privately on the build you send us, show you what failed and how to reproduce it, and retest once you have fixed it. The suite is the one the public bench uses; we do not add custom tests.
Who it is for
Engineering and security leads at companies that ship coding agents and want a clean run before a release is public and scored.
What you get for each run
- Results for every check in the suite against your named build.
- For each finding: what happened, guidance on fixing it, and a script that reproduces it.
- A signed test record of which checks passed on that build on that date.
- Full results within 5 working days of receiving the build.
- Support by ticket, with a first response within 6 business hours (UK, Monday to Friday).
How it works
- You send us the build.
- We run the full suite and return results within 5 working days.
- You get fix guidance and a reproduction script for each finding.
- You fix it. We advise; we do not write or change your code.
- We retest the fixed build and issue the signed test record.
Prices
| Option | Price | What it covers |
|---|---|---|
| One build, no commitment | £10,000 | One run plus one retest of that build |
| One month, no commitment | £10,000 | 2 runs of any builds that month, cancel any time |
| 12 months, billed monthly | £8,000 a month | 2 runs a month |
| 12 months, paid up front | £6,500 a month (£78,000) | 2 runs a month |
- On monthly plans a retest counts as a run.
- A run beyond your allowance is charged at the one-build price.
- Prices include any VAT due. The fee does not depend on the result.
New checks
We add checks every quarter. Each batch is announced publicly to all vendors 30 days before it counts in public scores. Subscribers are tested on new checks during that window at no extra cost.
Independence
- Public results come from the shipped release, on the bench's own schedule, whether or not a vendor uses Harness Watch.
- Payment never buys a preview, delay, edit, retest or veto of a public result.
- We share how to reproduce each finding. We never share the full check catalogue, the test rig or the tools that generate the tests.
- We publish the names of vendors who use Harness Watch, never their private results. Using Harness Watch is not an endorsement.
- Our free security disclosures continue unchanged.
What it is not
The signed test record is a test result, not a sign-off, a certification or a guarantee of security. It records which checks passed on one named build on one date, and nobody should rely on it as a statement that software is secure. Our liability is limited to the fees you paid in the previous 12 months. Harness Watch is not consultancy, and we do not build tests to order.
How to start
Email sales@agenticthinking.uk with the option you want and when your build will be ready. We confirm the start date, you send the build, and the 5 working days start when we receive it.