150 AI Agents Rated: What the Data Reveals

The first 150 ratings give a more usable view of the market than the early 50-agent sample. All 9 Kanon profiles are now represented at meaningful scale, with Operations Agents accounting for 29 records, Knowledge & Research for 25, and Orchestrator Agents for 24. The catalogue is no longer a narrow test set. It is broad enough to expose recurring structural patterns.

The visible grade distribution is concentrated in the upper-middle rather than at the extremes: 59 agents sit in the B band and 49 in the A band, while 19 land in C, 16 in D, and 7 in E. That is a more mature shape than the earlier database, but it is still not a premium market by default. The largest share of public records clears a credible diligence threshold without yet reaching the strongest evidence tier.

Profile averages reinforce that point. Orchestrator Agents currently post the highest average total score at 73.1, with Operations Agents close behind at 72.8. Financial Agents sit lowest at 56.1. This is not a claim that one class of system is inherently better than another. It is a statement about what the public record currently proves. Workflow and orchestration products tend to expose operating detail more consistently than finance-adjacent products, where trust thresholds are higher and published evidence is often thinner.

Security is the lowest average dimension across the full catalogue at 58.1. The next two lowest dimensions, Reputation and Transparency, are materially higher at 64.1 and 64.3. Stability and Coherence sit far above that at 75.9 and 77.4. The industry's weakest point is therefore not basic product continuity. It is the quality of public security evidence.

That gap matters because Kanon does not let high-level polish erase a weak security floor. Under Taxonomy v1.0, the Security cap applies in profiles where Security is not already the dominant weight. In the current snapshot, it has downgraded 14 agents. Eleven were held to a total-score ceiling of 50 because their Security score fell below 30. Three more were held to 70 because their Security score remained below 50.

The cap is disabled for Financial and Orchestrator profiles because Security already carries 45% and 40% of the weighting there. For the remaining cap-enabled profiles, 14 of 115 active agents are currently being constrained by the mechanism. That is a useful discipline for buyers: a strong interface, a coherent workflow, or a recognizable brand should not override an insufficient security floor.

The practical reading is straightforward. First, compare agents inside the right functional profile before comparing them across the entire catalogue. A B-band Financial Agent and a B-band Customer Support Agent are not solving the same trust problem, and their evidence burdens are different.

Second, look at Security before you look at surface-level completeness. The lowest average dimension in the market is also the one dimension that can still cap the total score. That makes it the cleanest test of whether a public record is ready for serious procurement review.

Third, treat public ratings as an entry filter rather than a final purchasing decision. The catalogue is now broad enough to support structured shortlisting. Final selection should still depend on profile fit, disclosed controls, and the quality of evidence attached to the public record.