How we work.

What we are

Agentic Thinking is an independent AI agent research lab in the UK. Our work has three parts.

Independence

An evaluation is paid for by the builder of the tool we examine. These rules keep the result independent of that payment.

• The fee is fixed and paid up front. It is agreed before work starts and does not depend on the result.
• Findings are not softened. You can correct a factual error; you cannot change a finding. Whether and how a report is published is agreed before work starts.
• No endorsements, badges or seals. A result is a record of what we tested, at one commit, not a recommendation.
• We never evaluate our own software: HookBus, AgentHook or AgentAuditor.
• Prior work is declared. If we have advised on a product's design, any later evaluation of it says so on its first page.
• No bug bounty work. We do not do bug bounty work and do not take bug bounty payouts.

How an evaluation is run

• Every evaluation starts from a fixed commit, built from a fresh checkout.
• The code under evaluation runs in isolated containers with no network access.
• Every evaluation includes negative controls: deliberate faults such as tampered evidence, replayed authorisations and bypass attempts. A control that is not caught is reported as a finding.
• We run the evaluation at least twice and report only verdicts the second run reproduces.
• Every finding cites an evidence file, and every evidence file is listed with its digest, so the pack can be checked after delivery.

See a sample evaluation pack.

Your code and evidence

• Used only for your evaluation. We do not share your code or evidence with anyone else, and we do not use it to train models.
• Run on infrastructure we control, in isolated environments, with no network access for the code under test.
• AI assistance is disclosed. AI agents help us run and review evaluations. Before work starts we tell you which model providers, if any, will see your code, and you can restrict this.
• You see findings first. Where a report is to be published, we send you the full pack 10 working days beforehand, for factual corrections only.
• Retention. We keep the delivered pack and its evidence for 12 months after delivery, so that a re-test can be compared with it, then delete our copies of your code and evidence unless you ask us to keep them longer.

How the work is done

The lab is one researcher directing teams of AI agents. The agents help build test rigs, run tests and review results.

The researcher checks every published result and every finding sent to a client or vendor before it goes out. AI review is used in addition to human review, not instead of it.

Declared interests

We research AI agents, and we also build software in this area. So that readers can weigh our work, we declare:

• We steward the AgentHook standard and maintain HookBus. Both are Apache 2.0. We have committed to transferring the AgentHook standard to an independent foundation.
• Agentic Thinking has developed AI agent governance and workflow software: AgentProtect, AgenticStudio and HookBus Agent. None of it is sold or currently offered, and we do not market it through our research.
• AgentAuditor, the recorder used in our incident research, is ours and is not yet published.
• Patent applications by the founder, Leo Ruocco: HookBus GB2608069.7 and CRE GB2604445.3 (pending, to be assigned to the company).
• We take on paid evaluations of agent runtimes and governance tools, under the rules above.
• We also run AgenticBench, a benchmark of AI coding agents, under its own published rules. It takes no money from the agents it tests, and those agents are not eligible for paid evaluations.

Report a security issue

If you find a vulnerability in HookBus, the AgentHook specification or fixture, any hookbus-publisher-* repository, or this website, email security@agenticthinking.uk. Please do not open a public issue first.

• We acknowledge every report within 5 working days.
• We agree a disclosure timeline with you, normally no more than 90 days.
• We credit you when a fix ships, unless you ask us not to.
• We will not take legal action against research done in good faith, within this scope, that avoids harm to users and data.

Our contact details are also published at /.well-known/security.txt.

Privacy

Agentic Thinking Ltd (company no. 17152930, Office 9, Sycamore Trading Estate, Squires Gate, Blackpool FY4 3RL) is the controller of the personal data described here. Contact: leo@agenticthinking.uk.

• This website sets no cookies and uses no analytics or tracking. Our web server keeps standard access logs for security.
• Emails and enquiries. We keep what you send us to reply and to do the work you ask for. Lawful basis: our legitimate interest in responding, or a contract with you. We keep correspondence for up to 2 years after our last contact, longer only where the law requires.
• Evaluation clients. Contact details and contract records are kept for as long as UK tax and company law requires. Your code and evidence follow the retention rule above.
• Research. Incident research uses public sources. Our tests use our own machines and accounts, generated data and fake secrets, not real customer or personal data.
• Our open-source code is self-hosted by its users. HookBus sends nothing to us.
• Service providers. Our email and web hosting providers process data on our behalf. We do not sell personal data.

You can ask for a copy of your data, or ask us to correct or delete it, by emailing us. If you are unhappy with how we handle it, you can complain to the Information Commissioner's Office.

Certifications and testing

Agentic Thinking does not hold third-party security or compliance certifications, and no independent penetration test of our code has been completed. Our open-source code can be reviewed in full on GitHub.

Contact

• The lab: leo@agenticthinking.uk
• Evaluations: partnerships@agenticthinking.uk
• Security issues: security@agenticthinking.uk