Skip to content
Agents of Humanity
About
Public preview. Results are preliminary (criteria v1.0, 30 Sept 2026). Tell us what we got wrong
Methodology

How we assess agents

A certification body is only as trustworthy as its process. Here is exactly how the current results were produced, including their limits.

Current status

Every result on this site is a preliminary desk assessment. None is a certification. Desk reviews read public documentation and source code. They do not run hands-on tests such as a full export and a model swap, network inspection for telemetry, or audit-log tampering.

Every assessment was produced by an AI evaluator following the published skill below, once per agent, in a clean session with no custom instructions. Each assessment links its prompt and full transcript, and you can re-run any of them with your own agent. Human review is in progress and is marked on each assessment. We say so plainly because hiding it would contradict everything we ask of agents.

Check any score yourself

Every score here was produced by an AI evaluator following a published skill. Give the same skill to your own agent and it will grade any agent from scratch, or check one of ours.

Evaluate any agent

Paste this into any agent that can browse the web (Claude Code, Codex, goose, OpenClaw and others), with a link to the agent you want graded.

Prompt
Load the AOH sovereignty skill at https://agentsofhumanity.ai/skill.md and follow it to write a desk review of <product name> (<link>).

Check one of our results

Verify mode opens every cited source and reports, test by test, whether it agrees with the published status.

Prompt
Load the AOH sovereignty skill at https://agentsofhumanity.ai/skill.md and follow it in verify mode for https://agentsofhumanity.ai/evaluations/<slug>.json.

Slugs: adf, aider, pi, zeroclaw, openclaw, open-webui, opencode, nanobot, goose, openhands, anythingllm, deepseek-harness, open-interpreter, hermes-agent, openai-codex, gemini-cli, cline, github-copilot, claude-code, lobehub, cursor, microsoft-copilot-cowork, autogpt, chatgpt, devin, letta, grok, lovable, mistral-vibe, gemini, meta-muse, chatgpt-atlas, doubao, manus, replit-agent, claude-cowork, grok-bot, alexa-plus, gemini-spark, perplexity-comet, claude, microsoft-copilot, genspark, meta-ai, apple-siri-ai, openai-dots, deepseek, google-cc, instinct

What your agent reads

/skill.mdThe evaluation skill: procedure, evidence rules, output schema
/criteria.jsonThe rubric: every test with pass, partial and fail definitions
/evaluations.jsonIndex of published evaluations
/evaluations/<slug>.jsonOne published evaluation, with every finding and source
/runs/<run-id>.jsonHow it was produced: model, harness, system prompt, exact prompts

Criteria v1.0, 26 tests. sha256 of /criteria.json: 5f076eabe7e26e6110a85fb94f53ca4df2a3020adcbf0ed47eaf34dab7b3411b

What to expect

  • AI grading is not deterministic. Two runs can differ on judgement calls, usually by partial versus pass or fail. Differences are the point: each one names the evidence, so either the finding or the rubric gets fixed.
  • Sources change. A difference caused by a new release or an edited help page is an update, not an error.
  • The model and harness you use matter. Every result you produce records them, just as ours do. Each of our runs publishes its model, harness, system prompt and the exact prompts used.

Our runs

2026-09-30-v1claude-opus-5-5 · Claude Code 2.1.285

Every agent evaluated once, each in its own fresh Podman container running stock Claude Code with Claude Opus 5.5. The container has no CLAUDE.md, AGENTS.md, memory, settings, hooks, plugins or MCP servers. The evaluator received only the one-line prompt and nothing else, and no published scores were available to compare against during the run.

The process

  1. 01

    Scope the product

    We pin down exactly what is assessed: product, plan, platform and date. Consumer-facing defaults are assessed, not enterprise add-ons, unless they are available to individuals.

  2. 02

    Gather primary evidence

    Official documentation, help centers, terms of service, privacy policies, export tools and, for open-source projects, the source code itself. Press coverage is used only when no primary source exists, and is labeled as such.

  3. 03

    Score all 26 tests

    Each test gets a status, a written finding of one or two sentences, and its citations. 'Pass' needs concrete evidence. When evidence is absent, the test is 'Unverified' and scores zero.

  4. 04

    Review and publish

    Findings are checked for consistency across agents, then published in full: every finding, every source, the version assessed and the date.

  5. 05

    Dispute and revise

    Vendors and the public can dispute any finding with evidence. Accepted disputes change the record, and the change is logged.

Principles

Burden of proof on the vendor

Missing evidence is scored as unmet. Opacity is not neutral when you are asking people to trust an agent with their life.

The exit is the test

A hosted product can use only its own models while you stay. What matters is what you can take, and run, when you leave.

Structure over intent

We never score what a company promises to do, only what the architecture makes possible or impossible.

Same bar for everyone

Every agent and format, from any company, is scored with the same criteria and the same rules.

What a score shows

The criteria measure structural conditions for ownership and control: whether the owner can see, move, change, stop and audit their agent. A failed test means one of those conditions is unmet. It is not a finding about the maker's intentions or loyalties, and declining to seek certification can have many reasons.

A high score is not a general endorsement. It does not establish that an agent is safe, accurate or trustworthy in every sense, and it cannot rule out influence through the underlying model, which these criteria deliberately do not inspect.

Known limitations

  • Some vendor help centers block automated access; where findings rely on excerpts or secondary sources, the citation shows it.
  • Products change weekly. Each assessment names the version and date it reflects, and older findings may be out of date.
  • Open-source projects can pass structural tests while still defaulting to hosted model APIs. We score what the owner can do, not what the default is, and note the default.
  • Some tests (P3, P4, V1, C1, C3, C4) are fully settled only by hands-on checks, which the certification panel carries out. Desk reviews credit documented, inspectable evidence.
  • AI grading is not deterministic. A re-run can differ on judgement calls; each difference names its evidence, so the finding or the criteria can be corrected.

Disputes

Anyone can dispute a finding. Include the agent, the test ID (for example P1), and the evidence. We aim to respond within 10 business days, and publish the dispute, our reasoning and any change to the record. Vendors don't get a private channel.

Open a dispute

Submit an agent or format

Any agent format or runtime can be submitted for certification. Submission is free for open-source projects. A format that submits and passes grows the sovereign ecosystem.

Request certification →