What actually changed?
Three capabilities stopped being demos and became table stakes at roughly the same time.
Planning — the model decomposes a goal into steps instead of answering a prompt. Tool use — it reaches outside its own context into your CRM, your repo, your browser. Execution — it acts, then checks its own output, then acts again.
The platform launches tell the story. OpenAI shipped background computer use on the desktop in April. Anthropic's approach is portable — screenshot, mouse, keyboard, no OS dependency. Google anchored to the browser through Mariner and pushed the Agent2Agent protocol into production at Cloud Next. Three different architectural bets, one shared assumption: the product is the doing, not the talking.
Underneath, the plumbing standardized faster than anyone expected. MCP went from an Anthropic side-protocol in late 2024 to roughly 41% of surveyed software organizations running servers in production, with over 110 million monthly downloads. When the integration layer commoditizes, the differentiation moves up — to workflow ownership — and down — to distribution. It does not stay in the model.
When "AI feature" meant a chat box bolted onto an app, you could charge for seats. When the AI closes the ticket, seats stop making sense.
The tell is in the pricing page
This is the part founders should stare at.
Per-seat pricing is structurally broken for agents: the better the agent works, the fewer humans the buyer needs, so the vendor gets paid to under-deliver. The market has already noticed. Seat-based pricing fell from 21% to 15% of SaaS companies in twelve months while hybrid models jumped from 27% to 41%. Support agents now price around $0.50–$2.00 per resolution, with escalations free.
Capital followed. Average agent-startup rounds roughly doubled — from about $82M in H1 2025 to around $155M by early 2026 — and by July 2026 the majority of deals are Series B+ at real revenue traction. That's not a hype curve. That's investors underwriting a metered, usage-linked revenue line and pricing it like infrastructure.
Now the uncomfortable number.
Somewhere between 85% and 89% of enterprise agent pilots never reach production. Gartner expects over 40% of agentic projects to be cancelled or shelved by 2027.
I don't read that as a refutation. I read it as the ordinary shape of an infrastructure transition. The 11% that ship are reportedly returning ~171% ROI, and the failure causes are boring and fixable: unclear success criteria, no tool or data access, no evaluation coverage. Nobody fails because the model can't reason. They fail because nobody defined done.
That's my actual thesis: agentic AI is not blocked on intelligence. It's blocked on measurement.
What to do with this
If you're building: stop selling capability, start selling a completed outcome you can name and instrument. If you can't measure the outcome, you can't price it — and if you can't price it, you're still selling a feature.
If you're investing: the durable question is no longer "which model?" It's who owns the workflow, and can they prove the agent finished the job? Evaluation infrastructure is the boring moat of this cycle.
The seat is dying as a unit of software value. What replaces it — resolution, task, outcome — is the entire fight of the next 24 months.
So: what's the first job in your company you'd genuinely pay per completed task for, instead of per seat? That answer is your agent roadmap.
© 2026 Tech Produit. All rights reserved.