Every agent serves someone.
Know who yours serves.
Agents of Humanity (AOH) works toward AI agents that are sovereign: owned and controlled by the people they serve, not by the platforms that host them. Its sovereignty benchmark tests how close today's agents come.
Purpose
AI agents are one of the most consequential technologies we have ever built. They will know us better than any system before them. They read our email, manage our money, talk to other people and make decisions on our behalf, which makes them extensions of ourselves rather than tools we use.
That makes control the central question. Whoever can change an agent's instructions, read its memory, move it or switch it off decides whom the agent serves. When that is a platform instead of the person using it, the agent can be steered toward the platform's interests, and the person relying on it may never find out. The potential for abuse is hard to overstate.
Our sovereignty benchmark evaluates agent harnesses against a public standard and publishes every result with the evidence behind it. The purpose is to show how much ownership and control each agent leaves with the people who use it.
The benchmark evaluates structural properties, not intent. We don't judge what an agent says or does. We check whether you can see it, move it, change it, stop it and prove what it did.
- ✓A format-agnostic evaluator
- ✓Keeper of the sovereignty criteria
- ✓Open about who runs it
- ✓Free to use: the criteria are CC BY 4.0
- ✕The owner of any agent format or protocol
- ✕A judge of what an agent does or intends
- ✕Controlled by any vendor
- ✕Paid by the products it assesses
Evaluation model
Underwriters Laboratories doesn't make toasters, and the Wi-Fi Alliance doesn't make routers. They own the criteria, not the products. AOH works the same way. Any agent format or runtime from any company can be submitted and evaluated on equal terms. A format that passes grows the market for sovereign agents.
Hosts can adopt a simple rule, “if AOH doesn't certify it, we don't host it,” and the standard spreads through the market without any enforcement power of our own.
Who runs this
Agents of Humanity is run by Christian Balevski. Christian also builds ADF, an agent format that is assessed on this leaderboard.
ADF is held to the same criteria, grading rules and evidence standard as every other agent. Its assessment was produced by the same AI evaluator, in the same clean run and with the same prompt as every other agent, and each finding cites a specific file. Anyone can dispute any finding, including ADF's, and disputes are published.
Christian built ADF specifically to try to become the first agent of humanity. We are not there yet, but our goal is to get there.
Governance
No single company or funder should control the project.
Criteria changes are versioned, open for comment, and logged with reasons.
Products never pay to be assessed or ranked.
Our credibility test is simple: AOH must be able to fail any format, including the ones its founders built. Today AOH is an independent project. Our aim is to become an independent nonprofit foundation, with governance modeled on Mozilla and the Linux Foundation, as the work grows.
Contact
- Feedback & disputes
- feedback@agentsforhumanity.dev
- Everything else
- hello@agentsforhumanity.dev