Why a catalogue, and not another list
Most agent lists give you a product name and a slogan, and that doesn’t help you choose. To decide, you need three things: what you have to hand over, what you get back, and who checks it. So The Agent Catalogue writes every listing as a contract. Each one has the inputs, the outputs, how the work runs, who reviews it, how hard it is to set up, and the price with the date I checked it.
An agent, for the catalogue, is a product you can sign up for or install. It takes a defined input, does research, drafting, analysis or actions on its own, and returns a defined output. Model APIs on their own, frameworks without a product, general chat assistants and prompt packs stay out.
281 listings sit across 14 categories. 36 are workflows I run myself through Guiding Ventures. The other 245 are third-party products, and the numbers below come from those. The catalogue opened on September 23, 2026 with 242 listings. In October I added 39 more, all launched between July and October 2026. Across all 281, marketing (30), sales (27) and coding (27) are the biggest categories.
What agents actually do
Almost every agent turns context you already have into a finished piece of work. A listing names 3.2 inputs and 3.2 outputs on average. The inputs are mostly things to connect: a code repository, a CRM, a knowledge base, your policies. The outputs are finished items: a merged pull request, a resolved ticket, a booked meeting.
| Agent | You give it | You get back | Listed price (Sep/Oct 2026) |
|---|---|---|---|
| Fin | Policies, brand tone, customer data, incoming conversations | Resolved conversations, refunds, account updates, handoffs | $0.99 per outcome |
| Agentforce Service Agent | Topics, instructions, actions, CRM and knowledge base | Resolved cases, password resets, handoffs | $2 per conversation |
| Devin | A ticket from Linear, Slack or GitHub, the repo, review feedback | Pull requests, bug fixes, migrations, docs | From $20 a month |
| ThunderPhone | A plain description of the agent, a phone line, documents, a contact list | Handled inbound and outbound calls, transfers, call scores | 2¢ per minute |
| Aider | A local git repo, a request, your own model API key | Edited files and commits | Free and open source |
| Harvey | Documents, legal questions, contracts | Sourced analysis, review findings, drafts | No public price |
Writing it down like this is the fastest way to compare. Two “AI SDRs” can look the same on their homepages and turn out to be very different once you list what each one needs from you. If you can’t say what you’d hand over and what you’d accept as finished, you aren’t ready to buy an agent yet. That’s also the first question I ask when I build agents for clients.
Half the agents don’t publish a price
117 of the 245 third-party agents (48%) show no price at all. It’s getting worse, not better: of the 39 agents launched since July 2026, only 7 have a public price. It also varies a lot by category. All 10 content agents publish a price, against 4 of 16 in HR and recruiting, 4 of 13 in automation, and 5 of 13 agent platforms.
Published prices can be messy too. When I checked in September, 27 pricing pages were unclear. One voice vendor’s homepage said “pay as you go”, but its pricing page showed only an Enterprise plan at $30,000 a year. One support vendor had two pricing pages that disagreed. Several prices were for the platform the agent lives in, not for the agent itself.
Here is how the 128 published prices break down. The grouping is mine, made by reading each price.
| How the agent is priced | Agents |
|---|---|
| Flat monthly fee | 94 |
| Per seat or user | 19 |
| Per unit of use (conversation, resolution, lead, credit) | 9 |
| Per minute (4 voice agents) | 4 |
| Other (one yearly plan, one per outcome) | 2 |
Paying for use instead of a flat fee is what should set agents apart from ordinary software, and it’s still rare: 9 of 128, plus the 4 voice agents that charge by the minute. It shows up in support, voice and sales, where the unit is easy to count: a resolved ticket, a minute on the phone, a qualified lead.
In the 2025 article I guessed that agents would soon compare vendors for us. An agent can’t compare a price that isn’t there. Once buyers send agents to build their shortlists, a hidden price stops being a sales tactic. It becomes a reason to be left off the list.
Who checks the work depends on the job
Vendors let agents run alone where a mistake is cheap and shows up quickly. Where a mistake costs money or ships into production, every coding, research, and finance and legal agent in the catalogue builds in a human review step, either required or optional.
| Category | Agents with no human review |
|---|---|
| Voice | 17 of 18 |
| Customer support | 16 of 20 |
| Coding | 0 of 27 |
| Research | 0 of 14 |
| Finance and legal | 0 of 15 |
A support answer that misses can hand off to a person. A wrong call transfer costs a minute. A bad migration, a flawed contract review or a research summary someone acts on costs far more, so those agents build in a review step. In coding, 9 of 27 require it and the rest offer it. In research, 10 of 14 agents require a review and the other 4 make it optional. 8 of 10 content agents require one too.
In 2025 I asked how we would learn to trust agents. The catalogue doesn’t show models making the difference. What the agents that run on their own have in common is a narrow job and a clear review step. When you look at an agent, ask where that step sits and who does it. If the answer is “you”, count that time as part of the price.
The listings go stale in weeks
A single re-check in October meant updating 16 listings. One agent had a new owner, another price dropped by more than half, and several descriptions had changed. I also added 39 agents that launched between July and October 2026. A directory of agents works less like a book and more like a price feed.
That’s why every price carries the date I checked it, and any price older than 90 days gets opened again, then updated or removed. Links get the same treatment: on day one, 337 of 345 links answered, and the rest were fixed or removed.
How the catalogue stays honest
- Facts come only from the vendor’s own pages, opened the day the listing is written. The source URL and date are stored with each listing.
- If a fact isn’t published, the field stays empty. No guessed prices, no “starting at” from a sales call.
- A script rejects sales words like “revolutionary”, “seamless” and “game-changing”, so listings describe instead of sell.
- Vendors can’t pay to rank. I pick the head-to-head comparisons myself instead of generating a page for every pair.
- Third-party listings are marked unverified until I’ve given each one a second look, and their links carry nofollow until then. That second pass is the next job on the list.
What held up from 2025, and what didn’t
The agent-first web held up, but less dramatically than I wrote it. Websites still matter. But the useful part of a vendor’s page is now the part a machine can read: what goes in, what comes out, what it costs. The catalogue stores each listing as structured data for exactly that reason.
Trust came from scope. The agents that run on their own are the ones with the narrowest jobs.
Agents buying for us is closer, but stuck on prices. The pieces exist. Half the agents in the catalogue still hide the one number an agent would need to compare offers.
What I underestimated was how narrow the useful agents are. Most listed products do one job with a clear input and output. General “do anything” agent platforms are 13 of 281 listings.
How to pick an agent
- Write down what you’d hand over. If you can’t, start there.
- Write down the output you’d accept as finished.
- Find out where review happens and who does it.
- Work out the price per finished piece of work, even if the vendor charges per seat.
- Check again in 90 days. The agent, its price and even its owner may have changed.
You can browse all 281 listings at agentcatalogue.com and filter by what you can give and what you need back. If nothing fits and you’d rather have one built around your own data, that’s what I do.
Numbers come from the catalogue’s listing files on October 21, 2026. Prices are as each vendor published them on the date stored in the listing. The price groupings are my own reading of those prices.