The agent economy is starting now. What does that mean?

What changed, in one paragraph

For most of the last decade, “AI in the business” meant a SaaS tool with a chatbot bolted on. That stopped being true in 2025. Models got cheap enough, reliable enough, and tool-using enough that you can now build software that does the work. The close, the triage, the sourcing, the reporting. Not helping a person do it slightly faster. Doing it. The companies we’ve seen wire this in first are pulling forward a kind of operating leverage the rest of the market will spend the next decade catching up to.

Why it’s happening this year

Three curves crossed in the last twelve months and none of them are reversing. Frontier model cost per task dropped about an order of magnitude. Tool-use and long-horizon reasoning quietly became reliable enough for production. The integration layer (every API your business runs on) became something an agent can navigate without a human writing the glue. None of these alone is a story. Together they are.

If you’re past the founder-does-everything stage and you have a real team running a real revenue line, this is the quarter when the cost of not building the stack starts to compound. A year from now, the gap between you and the company that started in Q2 2026 won’t be one of strategy. It’ll be one of installed surface area.

Three things you should know

From the engagements we’ve run, every business ends up needing the same three pieces. Not all at once, not in any particular order, but eventually all three.

1. A finance agent that operates like an institutional investor

Every company we’ve walked into runs finance the way a startup does. Month-end close. Quarterly board meeting. A CFO doing AP/AR by hand on Sunday nights. The ones we’ve helped pull ahead now run finance the way a portfolio company at a serious PE fund does: cash position read daily, variance against plan flagged inside 24 hours, the investor update drafted automatically against last month’s.

It isn’t glamorous and it isn’t the part of an AI pitch most firms lead with. It’s also the part that funds everything else. Once the finance layer is continuous instead of monthly, every other decision in the business gets sharper.

2. An internal agent stack. The team you didn’t hire.

This is the piece most people picture when they hear “AI agents.” It’s the work nobody on the team enjoys but that has to happen: pipeline hygiene, vendor renewals, the support inbox, scheduling, the second draft of everything. It’s now squarely in the zone of what an agent can do, not assist with. The agent lives in Slack, in the CRM, in the help desk, in the tools the team already uses, and quietly absorbs about a third of the week.

The shortcut is to buy an off-the-shelf SaaS tool that promises this. That gets you a slightly better inbox. The leverage we’ve seen comes from agents that know the business: the customers, the accounts, the products, the tone. That requires building for you, not buying from a catalog.

3. An external layer. Your business made legible to other agents.

This is the piece most teams haven’t started on, which is exactly why it matters. Inside twelve months, your customers will be sending agents to talk to your business. Not just their support agent talking to yours. Their procurement agent reading your pricing. Their research agent comparing your product. Their executive-assistant agent trying to book a meeting.

Companies with a clean, structured, predictable surface for that traffic (real APIs, structured data, a reliability layer they can audit) will get included in the network those agents are forming. The ones without it will be routed around the same way un-indexed sites got routed around by Google in 2003. The window to build this layer before it matters is open right now.

When the model picks the winner

Stack Overflow’s 2025 Developer Survey put Supabase at 6% adoption across all developers. Supabase’s own October 2025 numbers put it at 55% in the most recent Y Combinator batch.

A 9× skew between the cohort that builds with Claude Code and the rest of the market. Some of that’s product fit. Some of it is the loop: developers ask Claude what to use, Claude reaches for Supabase, the projects ship on Supabase, the next round of training data has more Supabase code, and the recommendation reinforces. Claude Code scaffolds projects with it by default. The Vibe Code Bench paper out of arXiv uses it as the substrate. Hacker News threads have been pointing at the same pattern for months.

You can’t fully separate the product from the loop anymore. Other categories are about to look the same way. Today it’s the database. In a year it might be the help-desk vendor your support agent recommends, or the supplier your customer’s procurement agent reaches out to.

The frontier models being trained right now ship in six to twelve months. They’re learning on what’s publicly indexable today — your docs, your API surface, your structured product data, your changelogs. If your business surface is legible to an agent right now, the next models will know how to interact with you. If it’s a Webflow page behind a “request a demo” button, they’ll learn you’re a thing customers email, not a thing agents transact with.

The reason to do this work this quarter isn’t operational. It’s that the next models are training on you whether you’ve prepared or not.

What “not ready” looks like

Concretely:

• Your finance close takes more than five business days.
• Your CFO’s most-used tool is Excel.
• More than a quarter of your support tickets are answerable from your help center but a human answers them anyway.
• Your CRM is a graveyard of half-filled fields and stale opportunities.
• The senior people in the company spend a meaningful fraction of their week on work a thoughtful 22-year-old could do.

None of these are crises on their own. Together they’re the operational signature of a business that’s about to be out-leveraged.

What “ready” looks like

A version of the same business where the close runs nightly, the CFO is doing strategy not data entry, support handles double the volume with the same headcount, the CRM is groomed continuously, and the senior people are working on the things only they can do. Same revenue, very different P&L. Same customers, very different experience. And, quietly, a company that other agents can do business with.

How we work

We don’t run twelve-month transformation programs. We don’t run workshops. We ship the first useful agent into production within thirty days, pick the one that pays for the next six, and build outward from there. The first three sessions are free because neither of us should be guessing whether this works on your business until we’ve looked at your numbers together.

If that sounds like the kind of thing your business should be doing this quarter, we’re open to two new engagements between now and the end of June.