The people who test them.
No single test picks a winner, and most rate the engine, not the car. The model is the engine. The agent is the car you'd actually buy. As of 7 Nov 2026, one test scores personal agents as whole products: PersonalAgentBench.
Which tests cover each agent
| Agent | Personal agentsPersonalAgentBench | Coding agentsAA Coding Agent Index | Self-hosted agentsWildClawBench | Office tasksGDPval-AA | Running a businessVending-Bench | Freelance workRemote Labor Index |
|---|---|---|---|---|---|---|
| Grok Bot | Tests this product | ?We can't say, because it doesn't show its model | Not tested | ?We can't say, because it doesn't show its model | ?We can't say, because it doesn't show its model | Not tested |
| Dots | Tests this product | Tests its model, not the product | Not tested | Tests its model, not the product | Tests its model, not the product | Tests its model, not the product |
| Muse | Tests this product | Tests its model, not the product | Not tested | Tests its model, not the product | Tests its model, not the product | Not tested |
| OpenClaw | Not tested | Not tested | Tests this product | Not tested | Not tested | Not tested |
| Hermes Agent | Not tested | Not tested | Tests this product | Not tested | Not tested | Not tested |
| Claude Code | Not tested | Tests this product | Tests this product | Tests its model, not the product | Tests its model, not the product | Tests this product |
| Codex | Not tested | Tests this product | Tests this product | Tests its model, not the product | Tests its model, not the product | Tests this product |
Tests this productTests its model, not the productNot tested?We can't say, because it doesn't show its model
Read 7 Nov 2026. We show who tests what, in our own words, link to the source for results, and say on each guide where a test disagrees with our pick.
What each test tells you
PersonalAgentBench
Scores personal agents on delegated tasks, and on whether they stayed within what you allowed.
The first test we found of the whole product. Its makers call the results early, from a small sample.
WildClawBench
Runs OpenClaw, Hermes Agent, Claude Code and Codex through the same 60 tasks.
One of the few tests of self-hosted agents themselves.
GDPval-AA
Ranks models on real work tasks from 44 occupations.
The closest thing to a test of everyday knowledge work.
AA Coding Agent Index
Tests coding agents directly, on results, cost and time.
One of the few tests of a product paired with a model, not only the model.
Arena
Shows which models people prefer in blind comparisons.
Real people, with no idea which model they are rating.
Vending-Bench
Tests whether a model can run a small simulated business for a year.
The closest test to running a small business.
Remote Labor Index
Measures how much real freelance work AI finishes to a professional standard.
Real paid projects, checked by people.
Epoch AI
Tracks benchmark results across models, and reviews whether the benchmarks hold up.
A good first stop for how much to trust a test.