> Source: https://theagentsindex.com/what-they-do/method

# How we grade what an agent does

The questions we ask, which plan we read, what each grade means and when a maker can change one.

**Method 1, used since 20 November 2026.**

## The questions

Each kind of work has its own short list of questions, chosen because agents doing that work answer them differently. For support: does it act inside your other systems, work in the help desk you already run, and test on past tickets before it goes live. For voice: is the minute price all-in, can it pass a live call to a person, and can it book a meeting during the call. For coding: does a task end as a pull request, can you pick the model, and is your code kept out of training. For sales: does it find prospects with the data included, answer replies and book meetings, and run month to month. A question every maker answers the same way is dropped. A grade says what the maker's page states. It doesn't test how well the agent does the job.

## The plan we read

We grade the plan we price at the standard volume. Where no plan has a price for it, we grade the cheapest paid plan the pricing page lists, and the row names it. A free plan that bills the work somewhere else isn't the plan you'd pay for, so we skip it.

## The pages we read

Questions about price and terms, such as what a plan bills for or whether it runs month to month, are read on the pricing page alone, twice, by two separate readers. Where the readers disagree, a third read settles it.

Questions about what the agent does, such as passing a call to a person or opening a pull request, rarely appear on a pricing page. For those we also read up to three of the maker's own product, docs or help pages, and the row links the page the answer came from. "Not stated" on these means none of those pages says. A maker with no public prices is graded on these questions from its product pages, and its price stays by quote.

## The grades

### On the plan

The priced plan's own page says it does this.

### Add-on

Sold on top of the priced plan, at a price the page prints.

### Higher plan only

The priced plan leaves it out and a dearer plan has it. The row names the plan.

### Not offered

The priced plan leaves it out and no plan above it adds it.

### Not stated

The page doesn't say either way. We count this as a finding: you'd have to ask before you buy.

### No plan details

The maker prints no plan details we can read. We treat it as unknown, not as a no.

### Under review

Our reads disagree, and the cell stays out of every count until another read settles it.

A grade marked **Partly** covers part of the question, and the row says which part is missing.

## Who is in the counts

Every agent we list for that kind of work: the ones buyers search for and the ones whose makers joined, read the same way. Joining never changes a grade.

## When a maker writes in

We change a grade only when a page anyone can open shows it: a pricing page, a plan sheet or the docs. A demo or an email doesn't count.

## Editions

This is the first edition, under method 1, read in November 2026\. Pricing pages are read again with each monthly edition, and every changed grade is dated. The cost per unit has [its own method](https://theagentsindex.com/methodology).

## How the method has changed

Each version of the method, its date and what it changed.

- **Method 1, 20 November 2026**: First edition: the questions that split agents doing the same work, 6 for support, 7 for voice, 7 for coding and 6 for sales, graded on the plan we price, from each maker's own pricing page.
