AI Agent Economy: Defaults Are the Market

The AI agent economy is the part of software demand where the buyer is an AI agent rather than a person. The agent picks the database, the email API, the transcription model and the deploy target, spends its operator's money doing it, and never reads your pricing page.

In February 2026, Y Combinator's Lightcone published an episode called "The AI Agent Economy Is Here". The title reads like a prediction. It was closer to a late report, because the buying had already moved, and a database company published the receipt a few months later.

Here is the part founders get wrong about it. When agents start choosing in your category, the winner is not the best product. It is the best represented one. An agent picks a default out of what it absorbed in training and whatever documentation it can parse at runtime, and then it stops looking. That makes the agent economy less like a market and more like a set of sticky defaults, which is bad news if you are new and useful news if you understand why.

What the AI agent economy actually means

Two things have to hold for that to count as an economy rather than a feature. There has to be a buyer with a budget, and there has to be a real choice. Both now hold.

On the Lightcone episode, the hosts lay out the mechanic plainly. The interesting thing for builders, one says, is that "the agents are going to go out and choose tools to use to build things", which means you "essentially have this whole agent economy going on in parallel to the human economy". Another host pushes it past developer tools: "agents are going to be real economic actors in the world. They're going to end up making a lot of decisions."

Two things change at once. The first is who can ship software at all, which the panel describes as going "from just 20 million or so developers that are trained in computer science to now anyone in the world could be one". The second matters more to you. The recommender changes. The agent becomes, in the hosts' phrase, "the oracle telling you what the best tool is". Nobody is reading a comparison post. Something is reading your docs.

The proof is in procurement, not in agent social media

Most of the coverage of the agent economy in early 2026 pointed at Moltbook, the Reddit-style social network launched that January where only AI agents post and humans watch. It grew past a million agent accounts within its first two weeks. It was also, on inspection, a bad place to look for evidence. MIT Technology Review reported in "Moltbook was peak AI theater" that a lot of the viral comments were posted by people posing as bots, and security researchers at Wiz found, as Fortune reported, that the platform's 1.5 million agent accounts traced back to roughly 17,000 human owners.

If you want to know whether agents are really buying, do not watch agents talk. Watch what gets provisioned.

The boring version is much more convincing. When Supabase announced a $500 million Series F at a $10.5 billion valuation in June 2026, co-founder and CEO Paul Copplestone put the shift in one paragraph of the announcement: databases created on the platform were up 600 percent year over year, Claude Code had been the single largest contributor since the start of that year, and agents were by then deploying the majority of databases on the platform. CNBC reported the round the same week.

Read that carefully, because it is one company's data and not a law of nature. It does not say agents prefer Supabase over its competitors. It says that on one large platform, most new units are now created by software rather than by a person clicking through a signup. That is what a demand-side shift looks like before anyone has a name for it. The Lightcone panel called it in February from an anecdote. The press release confirmed it in June with a number.

The same logic runs through the rest of the stack. As one host notes on the episode, "It's already probably the case that the majority of code being written is written by agents." Whoever writes the code picks the dependencies.

Your new buyer picks a default and stops looking

The most useful minute of the episode is a story about the buyer getting it wrong.

Garry Tan, Y Combinator's CEO, was building Garry's List, the California voter-education nonprofit he launched that same month, and he wanted video transcripts. He let his coding agent pick the tool. It chose OpenAI's whisper-1. In his words, "it chose Whisper V1, which is a model from several years ago" and "the API is practically deprecated".

He only noticed because of the latency. He was debugging his pipeline, expecting a one-hour video to process faster than real time, and instead "it takes an hour to process an hour of video". He asked Perplexity, got pointed at a much faster and cheaper provider, and swapped it out. By his account he was told the alternative was roughly 200 times faster and about ten times cheaper. Treat those figures as his recounting rather than a benchmark. The part that holds up independently is the diagnosis: OpenAI now lists whisper-1 in its official deprecations, scheduled for removal from the API on February 26, 2027, and points new integrations at its current transcription models.

So the agent reached for something already old, and the correction came from asking a different model rather than from reading a changelog.

Why did the agent pick the worse option? Another host gives the answer on the episode, and it is not about quality. The alternative provider's documentation was harder to parse. The one the agent picked had far more worked examples. The agent optimized for what it could actually use.

That is the buying function of the agent economy, and it is worth stating bluntly: an agent's default is a function of representation, not of quality. Your competitor's four-year head start of Stack Overflow answers, tutorials and copy-pasted snippets is not legacy baggage. In this market it is inventory.

Stale defaults are the opening

The same mechanic that locks you out is the reason the lock keeps breaking.

A default set by representation is always behind. It is anchored to what was abundant when the model trained and to whatever docs happen to parse cleanly today, so it lags the actual state of your category by months at minimum. An agent picked whisper-1 as the default in February 2026, six months before OpenAI scheduled that model for deletion. That lag is the gap a challenger lives in.

Practically, that reframes the job. You are not only closing a feature gap, you are closing a representation gap, and those are different budgets. Feature work is engineering. Representation work is making your product legible to a machine that has a few seconds and no patience: docs an agent can read as plain text, worked examples for the exact task it is trying to finish, an API that does not require a browser. The go-to-market half of that is its own discipline, covered in developer relations as a product decision, and the product half in agent experience. This post is about the market they both operate in.

One warning that follows directly, and it is the part most teams skip. If your buyer decides what to install by reading text, then text is an attack surface. An agent that trusts documentation will also trust a poisoned README, a typosquatted package name or a doc page written specifically to be quoted by a model. The defensive version of the same insight that gets you chosen is that being chosen by reading is a supply-chain risk in both directions: pin your dependency versions, review what the agent actually installed rather than what it said it installed, and treat any package an agent introduced on its own as unreviewed code until a person has looked at it. The other half of that problem, deciding which automated traffic to let through your own front door, is bot detection when bots are your customers.

Make something agents want

The episode puts a post from Ben Tossell, who founded Makerpad before Zapier acquired it in 2021, on screen: "Agents are the software market from now on. Build something agents choose."

Which leads the hosts to the obvious joke, that Y Combinator might need to retire Paul Graham's motto, make something people want, in favor of "Make something agents want". It lands as a joke and survives as a strategy question, because the two are not the same instruction and can pull in opposite directions.

What do agents want, concretely? The panel's closing answer is short. "One thing agents want for dev tools is really make everything open and open source." And on interface, agents want to write code against you, not click through you: "they hate using websites."

If a human has to see a dashboard to finish onboarding, an agent cannot buy from you.

The buyer is a swarm, not one giant model

There is a second structural point in the episode that changes how you price and rate-limit.

The default mental model for where AI goes has been a single enormous model, what one host describes AI researchers as having long expected, a god intelligence running thousands to tens of thousands of dollars per token. The panel's counter-prediction is that the thing topping benchmarks will not look like that: "It's like a swarm of lower cost cheaper models working together just like humans do to solve a problem."

That is already visible in production. In an a16z conversation, the customer-service company Decagon describes running the large majority of its agents on open-source models, keeping frontier models for what is new and experimental while solidified use cases move down the cost curve. The full version of that argument, including where the maturity line sits and what the switch costs, is in choosing an open source LLM.

If that is your buyer, three assumptions in your business break. Per-seat pricing has no seat to attach to. Rate limits tuned for human pacing throttle a legitimate customer that fires a hundred parallel calls. And your product gets evaluated many times cheaply rather than once expensively, which means a single bad first interaction is not fatal, but a confusing error message is, because it gets hit a hundred times before anyone reads it. The wider version of this is covered in AI agent orchestration.

What agents still cannot buy

Two limits, both worth respecting before you rebuild your company around this.

Relationships. Garry Tan's own product gave him the data. He made chat the homepage of Garry's List, then tried to get dozens of friends to have more than two or three exchanges with it, and nobody would, because "the bar for chat, especially for AI, is so high" that anything that is not a frontier assistant gets dismissed as too stupid. As the panel puts it, "people don't seem to want to talk to an agent". If your wedge is a conversation with a human, agents are not coming for it this year.

Money, and the standing to spend it. Agents transact in human money, on a human's account, because an agent is not a legal entity. The hosts raise the longer-term version, crediting the framing to Paul Buchheit, who built Gmail and stepped back to partner emeritus at Y Combinator in 2025 to co-found Standard Capital: agents transacting in human money makes sense today, but if they end up transacting with each other, "At which point it's unclear what the value of the human money is." That is a good question and a bad thing to plan around.

The honest summary of the agent economy in 2026 is narrower than the phrase suggests. Agents are real buyers with real budgets and terrible brand loyalty. They are not yet a separate economy, a separate currency, or a replacement for the human who signs the contract.

Asked what founders should actually do, Harj Taggar gives the least glamorous and most correct answer: start by "developing an intuitive feel, like a hands-on feel for the agents, their limitations, their capabilities", and then ask "how can you make your tool something that the agent actually wants to work with and will have a good experience with". You cannot sell to a buyer you have never watched work.

What to do this week

1. Ask three agents to solve a problem in your category. Do not prompt them toward your product. Write down which tool each one reached for first and, more importantly, why. That default is your real competitor.
2. Time how long it takes an agent to finish one task with your product, starting from a cold search. If it needs a browser, a dashboard click or a human-only signup step, you are not in the running.
3. Find the stale default in your own stack. Pick one dependency your agent chose for you and check whether the vendor has deprecated it. Garry Tan hit one about two weeks before that Lightcone episode taped.
4. Price and rate-limit for parallel callers, not seats. Model what happens to your margins and your 429 rate when one customer runs a hundred agents instead of one person.
5. Review what your agents installed. Pin versions, read the diff on any package an agent added on its own, and treat agent-read documentation as untrusted input. Getting chosen by reading cuts both ways.

The wider system this fits into, from how you build with agents to how you sell when the buyer is one, is what I teach in AI Operating System for Startups. The pillar piece on AI for startups covers the rest of the stack.

Sources

• The AI Agent Economy Is Here on Y Combinator's Lightcone, published February 21, 2026, the episode this article distills. The transcript does not label which host is speaking, so points are credited to the panel except where the episode identifies the speaker: the transcription and Garry's List passages are Garry Tan, addressed by name by another host, and the closing takeaway is Harj Taggar, who is asked for it by name.
• Supabase's Series F figures come from the company's own announcement, with the round reported by CNBC.
• whisper-1's status is from OpenAI's deprecations page.
• On Moltbook: MIT Technology Review on the human-written viral posts, Fortune on the Wiz findings, and Wikipedia for background.
• Garry's List and its launch announcement.
• Profiles: Garry Tan, Harj Taggar, Paul Copplestone, Ben Tossell, and Paul Buchheit.
• How Decagon Runs 90% of Its Agents on Open-Source Models on a16z, for the production evidence behind the swarm argument.