Software for Agents: A New Market Nobody Is Counting

Here is a kind of purchase I make all the time now. I sign up for a service, enter a card and approve a budget. Then I never open it again. An AI agent installs it, runs it and is its only user. The sandbox it runs code in, the inbox it reads, the browser it drives and the wallet it pays from were all bought by me and used by it.

My thesis is simple. A new category of software is forming: software humans buy but never touch. It is real, it is growing fast, and no market report counts it yet. I went looking for the number and could not find one, so this post lays out what I could verify instead. I checked every source below on November 13, 2026.

What “software for agents” means

Most software agents use today was built for people. Agents drive GitHub, Vercel, Supabase and Stripe because those companies added a surface agents can reach, usually an MCP server. That is software agents merely use. It is useful, but the human is still the intended user.

Software for agents is different. The agent is the user. The product’s basic unit is an agent’s sandbox, inbox, browser session, memory or payment, and a human only provisions it and pays.

You can tell which is which by a few signals:

• Output for machines. Ramp’s CLI has an --agent flag for “machine-readable JSON output (default when piped)” and a separate --human flag for tables (Ramp CLI).
• No prompts. The tool runs without interaction and takes its credentials from the environment. Ramp has a --no-input flag and an agent login command for this (Ramp CLI).
• An identity for the agent. Stripe’s MCP server takes “agent API keys” with their own badge, and from December 2026 it stops accepting keys without the Agent tag (Stripe).
• A human approval step built in. Stripe marks tools such as refunds and send_money for human confirmation in the Dashboard (Stripe).
• An MCP server and docs written for agents. Vercel’s docs now have sections addressed to AI agents (Vercel).

Stripe and Ramp are human products growing agent features. The interesting question is what gets built when the agent is the only user from day one.

The landscape

Open source

The base layer is open. The Model Context Protocol is how agents reach tools, and in December 2025 Anthropic gave it to the Linux Foundation’s new Agentic AI Foundation, alongside Block’s goose and OpenAI’s AGENTS.md (Linux Foundation). Nobody publishes an official count of MCP servers, and the directories disagree by a factor of five: PulseMCP lists 21,552 (PulseMCP) and Glama lists 120,540 (Glama). Either way, it is tens of thousands.

On top of that sit open-source tools whose user is an agent: E2B gives agents sandboxes to run code, Browser Use lets them drive a browser, and Mem0 gives them memory. Agent workspaces are the newest group. I covered them in Chat Systems for AI Agents: Block’s Buzz gives each agent its own key, and Agent Relay lets coding agents join through a CLI.

Commercial

The pricing pages tell you who the user is. I checked nine products today.

Product	Sells	Priced by
E2B	Sandboxes for agent code	Plan plus per-second usage
Daytona	Sandboxes for agents	Per-second usage
Browserbase	Hosted browsers for agents	Plan plus browser-hours
AgentMail	Email inboxes for agents	Plan by number of inboxes
Mem0	Agent memory	Usage
Exa	Search for agents	Per request
Firecrawl	Scraping and search	Monthly credits
Composio	Tool connections for agents	Per human seat plus tool calls
Ando	Team chat with agents as members	Per human seat; agents never take a seat

Seven of the nine bill by what the agent consumes: seconds, hours, inboxes, requests. The two that bill per seat bill per human. None charges per agent. AgentMail sells inboxes; its seats are for the people who log in to its dashboard, not for the agents using the inboxes. In every case a person signs up and pays, and the agent does the using.

Y Combinator

YC is funding this directly. I checked YC’s own company pages and found 13 companies from 2025 and 2026 batches whose user is an agent. A few:

• Maritime (Fall 2026): managed cloud for AI agents.
• AgentMail (Summer 2025): email inboxes for agents.
• Cua (Spring 2025): a cloud desktop for every agent.
• Archal (Summer 2026): API sandboxes built for agents.
• Salus (Winter 2026): checks an agent’s actions before they run.
• Sponge (Winter 2026): financial infrastructure for the agent economy.

The others are Arga Labs, Chronicle Labs, Kernel, Metorial, Notte, Orthogonal and Playgent. I counted a company only if YC’s page describes a product an agent installs, runs, runs on or calls while it works. Tools for the people who build or monitor agents did not count. That is a floor, not a total: YC’s directory search does not work without JavaScript, so I could not list every company.

Agents start to pay

The clearest sign that this is a market and not a curiosity is that the payment rails are being built for agents as the payer. Four launches in the first four months of 2026:

• February 10. Stripe began letting developers charge agents for API, MCP and HTTP requests in USDC on Base through the x402 protocol, as a preview (The Block, Stripe).
• March 18. Stripe and Tempo published the Machine Payments Protocol, which Stripe calls an open standard “for agents to pay” (Stripe).
• April 2. The Linux Foundation launched the x402 Foundation, with x402 contributed by Coinbase (Linux Foundation).
• April 29. Stripe announced that businesses can grant agents the ability to pay with Link’s agent wallet, with spending approvals (Stripe).

The shape matches the rest of the category. A human sets the budget and the approvals; the agent spends.

How big is it?

Nobody knows, and that is the point of the title. I read the market reports that come up first for AI agents. Every one of them sizes the agents themselves, or the platforms companies use to manage agents. None has a segment for tools that agents install, use or pay for. So the best I can do is proxies, with their definitions:

• BCC Research sizes AI agents at $8.0 billion in 2025, growing to $48.3 billion by 2030, a 43.3% annual growth rate (BCC Research).
• MarkNtel Advisors sizes AI agents at $7.6 billion in 2025, growing to $110.5 billion by 2032, 47.13% a year from 2026 (MarkNtel). Its own September 2025 edition said $42.7 billion by 2030 (PR Newswire), which tells you how loose these numbers are.
• Technavio puts enterprise agent orchestration platforms at 30.1% annual growth, adding $7.53 billion from 2026 to 2030 (Technavio).

Read these for direction, not size. They agree that agents are growing at 30% to 47% a year. If agents grow that fast, the things agents use grow with them, because every agent needs somewhere to run, something to browse with, an inbox, memory and a way to pay. But that is my inference, not a measurement. The honest answer is that the market for software whose user is an agent is unmeasured.

What’s next

The pieces that exist today are infrastructure: compute, browsers, inboxes, memory, payments. The next layer is applications an agent picks up for itself. Once an agent can find a tool, install it from the command line, authenticate as itself and pay within a budget a person set, the human step shrinks to one approval. Stripe’s agent keys and human-confirmation badges already show what that approval looks like.

Disclosure: I’m building one of these, Botpenny.

What builders should do

If you are building software, ask who your user is. If it is an agent, build for it:

• Make every command work without prompts, with JSON output and clear exit codes.
• Give agents their own identity and credentials, separate from the human’s.
• Put the human where the human matters: approval and payment.
• Price by what the agent uses, not by seats. The agent is not a seat.
• Ship an MCP server, and write docs an agent can read.

The takeaway: there is now software whose only user is a machine, sold to people who never open it. It has its own pricing, its own payment rails and its own YC companies. Nobody is counting it yet, which is exactly why it is worth building for now.