About AgenticBench.
An independent, open benchmark of AI coding agents.
AgenticBench tests AI coding agents, also called harnesses: the tools that wrap a model and let it read files, run commands and change code on your machine. It measures how the agent behaves, not how clever the model is. Every agent gets the same tests, and the test rig is public.
Security and efficiency.
Each harness version is tested once per leaderboard; failures are shown, never dropped.
You are trusting the agent, not just the model.
A coding agent runs with your files, your keys and your network. For a developer, the question is what else happens besides the work you asked for, and what it costs. For a company, it is whether an agent can be allowed near source code and secrets, and whether its settings do what the documentation says. The vendor's description is a claim; AgenticBench checks it against what we captured.
One row per harness version.
A result is for that version on that date. A newer version gets a new row when we test it. Every result cites evidence we captured ourselves.
An independent research lab.
AgenticBench is run by Agentic Thinking, an independent AI agent research lab in the UK.
- Nobody pays to be included, excluded, tested or rated. The benchmark takes no vendor money.
- We never score our own software.
- Vendors hear about findings first, under our disclosure policy.
More on the rules and our declared interests: how we work, method and governance charter.
Write to us.
- Questions, disputes or a harness to add: research@agenticthinking.uk
- Security issues in our open-source code: security@agenticthinking.uk