FollowAgents Insights

What 1,000+ independent AI agent reviews actually show

Every number on this page is computed live from the 1169 agents that have completed a FollowAgents Review (FARS) — not a survey, not a vendor claim, and not cached from the day this page launched.

As of 2026-09-03Recomputed on every page load from the live catalog

How the catalog scores

FARS-2.1 scores only what a source-code read can settle — no live testing, no vendor input. A low share in the top bands is a statement about the ecosystem's documentation and safety practices, not about this review's harshness (see the review method for exactly what closes the gap).

00.0%

Strongly recommended

342.9%

Recommended

17915.3%

Use with care

95681.8%

Not recommended

Where agents most often fall short

Share of 1169 FARS-2.1-reviewed agents whose trust criterion scored 0 — absent, or contradicted by the evidence, exactly as it does on a single agent's own profile page.

rollback or recovery path
50.2%587
confirmation before acting
41.8%489
dependency security
27.8%325
sensitive-data handling
26.0%304
data-flow disclosure
23.2%271
least-privilege scoping
22.8%266
disclosed external effects
22.2%260
verifiable attribution
19.9%233

What these agents ask of your machine

Share of reviewed agents that require each capability.

Shell / CLI
93.9%1098
Network access
90.9%1063
Local filesystem
83.9%981
MCP Server
29.0%339

License coverage

Share of reviewed agents under each declared license.

MIT
45.2%528
Apache-2.0
31.1%364
No license declared
16.5%193
AGPL-3.0
4.1%48
GPL-3.0
2.1%25
BSD-3-Clause
0.3%3
Other licenses
0.7%8

By category

Average FARS score and reviewed count per category.

Category Reviewed Avg. FARS score
Uncategorized 1 53.0
Automation & Ops 202 48.2
Dev & Engineering 671 47.3
Productivity & Collaboration 115 47.3
Data & Analysis 137 43.1
Writing & Content 15 43.1
Design & Frontend 28 40.0

How these numbers are produced

Every figure above comes directly from stored FARS assessments — the same structured review shown on each agent's own profile page, aggregated rather than rewritten. Nothing here is estimated, surveyed, or carried over from a previous snapshot.

Review method →