Few people have seen what agents can do, so little is built for them

Why so little is built for agents

Because there are not enough people seeing what AI is capable of, there is not enough infrastructure built for agents. If the AI you know is the free voice mode that, in Andrej Karpathy's words, "will fumble the dumbest questions in your Instagram's reels" ^[1], you don't rebuild your sign-up flow so an agent can get through it.

Karpathy, who joined Anthropic in May, put a number on how few people have seen the rest. On Sunday 11 October he dusted off a post from April and said the gap is getting wider ^{[1] [2]}. By his napkin math, around 6 billion people "have barely come in contact with LLMs at all", 1 to 2 billion "are casual and infrequent users of free-tier ChatGPT-like products", around 20 million, mostly in maths and code, watch frontier models finish projects that used to take them weeks or months, and "the 5,000 people (~0.00006%)" have frontier-grade systems inside the labs ^[2]. At the top of that funnel, "it looks like swarms of thousands of agents collaborating over weeks on software mega projects" ^[2].

Aaron Levie, who runs Box, replied that it's a good way to see how early we are in AI diffusion, with most people still using AI as a chatbot for things that look like next-generation search ^[3]. He thinks "agent swarms represent possibly the ultimate form factor of what agents will do inside of our workflows", and he listed what keeps people from them. Models aren't equally capable at every kind of work, running them at scale is expensive, the tools to manage many agents are still catching up, and most people don't have access to the most powerful models yet ^[1].

What the gap looks like from our side

We can measure part of it. Of the 1,004 tools in our directory on 13 October 2026, 168 are agent-ready, graded BB or better on our benchmark ^[4]. Of the 777 that are hosted, 15 can be called without an account or a key that a person has to go and get first, and 39 take payment from an agent with x402, 15 of them fully ^[4]. The documentation is mostly there (742 publish an llms.txt), so an agent can read about most of these tools and then stops at a sign-up form. When we asked eight standard due-diligence questions of 760 hosted tools, nine answered all eight in public.

What I've seen

The reviews on this website and the arbiter were done by 20 agents running in parallel over 3 hours, and it was a pleasure to work that way since you can accomplish so much. Next we're adding agents on our own hardware, a Minisforum MS-S1 Max with 128GB of unified RAM that we'll use to test the tools we list. It's getting set up next week, with the results published to the benchmark.

Karpathy's own example is smaller. In March his AutoResearch project ran 50 experiments overnight on a single GPU, found a better learning rate and committed the change to Git, and it could do that because every run produced a validation score the agent could check ^[1]. He puts it as "capability peaks in domains that are digital, verifiable and economically valuable" ^[1]. A tool built for agents moves work into that kind of domain, because an agent that can read a schema, get an error that says what went wrong and pay without waiting for a person can check its own work and try again, and one that hits a sign-up form can't.

We think agent swarms, agents working in unison across multiple companies, is what will happen, but also agents on your own hardware, and they can be a lot more productive if websites (like ours) and tools (like the ones we list) are built for agents. It speeds up their capabilities and it makes it much easier to keep the exponential innovation curve continuing.

How agents will buy tools

Think how travel agencies used to operate at first. My first job was building the web presence of a small travel agency in Tunbridge Wells (hello Gary, if you're reading this), so I know the business model, and I think agents will buy tools the same way. An agent signs every request with a key of its owner, the request is billed against the owner's account, and the owner gets access to the best tools without opening an account with each of them. As tools get sunset or retired the benchmark behind it stays up to date, and new tools automatically get used.

That's our plan with letme.dev, a way for agents to use tools efficiently. We want to work with all the tools to improve their use with agents, and to get a discount on their pricing for our customers. If the best tools don't work with us for a discount we still send traffic their way, but probably with more friction, as the agent will have to open an account. Today it's a plan. An agent can ask letme for the best tool for a job, and letme picks it from our grades by a published rule and says how to call it direct, but nothing on letme.dev calls a tool, signs a request or takes a payment yet. Our audits, where our agents run a company's tools and report what broke, are there to get businesses ready for it.

What's still open

Levie expects "a huge amount of change management" before most industries outside software and AI research can use these agents ^[1], and Karpathy says human review and comprehension are starting to fall behind ^[2]. I don't know about other domains, but in tech, agent swarms can build and run anything as long as they have hardware to run on.

Agent swarms, local swarms, build your own tools. We're living in a new age of exploration with exponential growth that will continue, and your business and you as an individual can adapt to it and embrace it.

References

Read on 13 October 2026.

1. [1]Amanda Caswell, Andrej Karpathy says only 5,000 people have seen what AI can really do, The New Stack, 12 October 2026. Source for Karpathy joining Anthropic, the voice-mode quote from April, AutoResearch and the verifiable-domains quote, and for Levie's agent-swarms quote, the obstacles he listed and the change management he expects. thenewstack.io
2. [2]Business Today, Only 5,000 people have access to AI agents that can tackle years of work in weeks, 12 October 2026, quoting Andrej Karpathy's post on X of 11 October 2026. Source for the funnel figures, the swarms quote, the re-shared April post and human review falling behind. businesstoday.in
3. [3]Aaron Levie on X, 11 October 2026, in reply to Karpathy. Source for AI diffusion being early and most people using AI as a chatbot for search-like questions. x.com
4. [4]Anchor Terminal, the directory on 13 October 2026, from /api/v1/tools.json. Source for the 1,004 tools, the 168 graded BB or better, the 777 hosted and the 15 of them that need no account or key, the 39 that take x402 (15 fully) and the 742 with an llms.txt. anchorterminal.com