// THE AGENT TEST · 30 PRODUCTS · PUBLIC EVIDENCE

Is it actually an agent?

Thirty AI products judged against five published criteria. Every verdict links to its evidence, corrections and community vote.

01

The admission test

versioned rules →

25 of 30 pass the Agent Test. We corrected ourselves 32 times getting there — 30 of them against our own thesis.

New here? Start with the plain-English answer to “is it actually an AI agent?” →

01LOOPSruns until done
02CHOOSESpicks its own tools
03ACTSchanges the world
04RECOVERSfixes its own errors
05UNSUPERVISEDnobody watching

4–5 · AGENT 2–3 · WRAPPER WITH AMBITION 0–1 · CRON JOB WITH EXTRA STEPS

Score one we missed →

  1. LOOPS
  2. CHOOSES
  3. ACTS
  4. RECOVERS
  5. UNSUPERVISED
03

Thirty product labels

opinion, evidence, correction path

Capability: how many of the five admission checks it passes · proof: how checkable our answer is · leash: tool calls between human turns, shown only when measured from a real trace.

ProductProofCapabilityOur verdictLeashThe people sayYour call

Verdicts are our opinion of the shipped product vs its marketing, scored on the test above. Disagree? That's what the buttons are for. Wrong facts? Tell us.

04

Where we are weakest

our evidence, not the crowd's opinion

This is the list we would attack first if we wanted to prove this site wrong. Community rankings — most disputed, biggest disagreement — appear here when there are enough votes to mean anything, and not before.

05

The record

the court → · rss

    Every verdict change, correction, and new judgment lands here with a date. Overturned calls stay on the record — that is the point of keeping one.

    05

    Submit a product

    we judge every submission by hand

    Sent to a private server-side review queue. Nothing is published automatically; a receipt appears here when the record is safely stored.