Most software you've ever used has the same fundamental shape. A human opens an app, clicks something, gets a result. The software is reactive — it waits to be told what to do, executes a narrow task, then waits again. Every productivity tool, CRM, ERP, and dashboard ever built shares this DNA.
Agentic AI breaks that pattern. For the first time, software can be assigned an outcome instead of a task. "Book me a flight to Austin under $400 next Thursday" replaces twelve clicks across three websites. "Reconcile last quarter's expenses and flag anything outside policy" replaces an analyst's afternoon. The software runs the workflow itself, makes decisions, recovers from errors, and reports back.
This is not a feature improvement. It's a new category of software, and it requires a new category of company to deliver it. The companies that win this layer won't be the ones bolting agents onto existing SaaS — they'll be the ones building from the agent outward.
Why SaaS pricing breaks
There's a reason the existing incumbents struggle here. SaaS business models are built around seat-based pricing, which assumes one human per license. Agentic systems collapse that assumption. When one agent does the work of fifty seats, the entire commercial model has to be rebuilt around outcomes — usage, success rates, hours saved, dollars collected, tickets resolved.
Incumbents won't cannibalize themselves to get there. A public SaaS company priced at 12x ARR cannot rationally swap its $50/seat/month line item for an outcome contract worth one-tenth as much in headline revenue, even if the underlying value to the customer is higher. New companies, with no installed base to defend, will. That asymmetry is the opening.
The three-layer stack
We see three layers forming in agentic AI, and the venture economics of each are radically different.
1. Foundation
The model labs and the infrastructure they require — compute, training data, alignment research, safety. The capital required is enormous, the winners are largely identified, and the venture math is brutal. We do not invest here.
2. Orchestration
The frameworks, memory layers, evaluation tools, vector stores, and runtime environments agents need to operate reliably. This layer is interesting, sometimes underpriced, and produces a few real businesses per cycle. It's also being absorbed into the model providers themselves at an accelerating rate. We invest here selectively, and only when the moat clearly does not depend on a model lab choosing not to build the same thing.
3. Application
Vertical agents that solve a specific business workflow end-to-end — billing for plumbers, intake for med-spas, lead qualification for staffing agencies, dispatch for HVAC. Boring on the slide, enormous in the wallet. This is where we focus.
Why the application layer wins
The application layer matters because it's where revenue actually compounds. Foundation models are commoditizing faster than anyone predicted — the gap between the leading frontier model and the third-best open-source model has narrowed every quarter for two years. Orchestration tools are being absorbed into the model providers themselves. But a vertical agent that handles every billing dispute for a regional plumbing chain — that's a defensible business with real customers, real switching costs, and real revenue.
Where defensibility actually comes from
- •Workflow integration depth — the agent is wired into the customer's calendar, payments, communications, and CRM. Ripping it out is a week of pain, not a Tuesday afternoon.
- •Proprietary workflow data — every interaction makes the agent better at this customer's specific edge cases. Generic models can't replicate that.
- •Outcome accountability — the agent is on the hook for revenue, not for clicks. Customers don't switch out a vendor that's collecting their money.
- •Distribution lock-in — the founder owns the trade association relationship, the franchise contract, or the channel that the next vendor would have to rebuild from scratch.
The historical precedent
Mobile took ten years to fully play out. Cloud took twelve. Agentic AI will move faster because the ingredients already exist — capital is liquid, talent has been pre-trained at the model labs and is now spilling out, and distribution to small businesses through search, social, and trade channels is dramatically more efficient than it was in 2007.
The companies being founded right now are the ones we'll all be using in 2030. The question isn't whether agentic is the next platform. It's whether you can spot the right founders before the consensus forms, and whether you can write the first check small enough to matter.
“Consensus is what gets priced. Conviction is what gets paid.”
We're a pre-seed fund, which means we are explicitly designed to be early to that conviction. By the time agentic AI is the obvious answer at every IC meeting, the prices will already reflect it. We'd rather be wrong about a few founders than late on the cycle.
