1/ "Most SaaS should become an MCP server."
@nicbstme's post making that case hit 175K views this weekend. I agree with half of it. The half I'd push on is the business model. 🧵
6/ My take: whoever holds the record of what the consumer approved (budget, merchants, limits) becomes the trust layer for agent payments.
Should that record live with the bank, the agent, or a network in between?
5/ Plaid seems to see it too. Next on its roadmap: a way for consumers to approve specific actions, and its Trust Index scoring agent activity in context.
Detection is step one. Permissioned actions are where the value is.
4/ But detection answers "is this an agent?" The bank's real question is "did my customer authorize this, for this amount?"
That's a mandate problem. It's the gap in Sardine's example from yesterday: a $1,400 laptop bought on a $1,000 budget.
3/ Why Plaid can do this: it already runs device and network intelligence on financial logins across hundreds of institutions. So it can tell an agent logging in as a bank customer from an agent shopping.
Few banks could build that view alone.
2/ How it works, per Plaid: thousands of device, behavior, automation and network signals flag agent sessions and the AI provider behind them.
Plaid says side-by-side tests on live traffic caught 2.3x more agent sessions than competitors, with no rise in false positives.
1/ Plaid just launched Agent Detection. It tells a bank whether a session is a human or an AI agent, and which agent.
Useful. But knowing which agent is at the login page doesn't tell you who sent it, or what it's allowed to do. 🧵
8/ I'm researching agentic payments at Stanford, and before that I ran an early-stage VC firm. I write about agent wallets, mandates and how agents will pay for software.
Follow @emmacui for more on agentic commerce.
6/ My bet: the winners price where the agent transacts. Metered per call or per outcome, paid from an agent wallet with spend limits the human set, permissions and audit built in.
That's where SaaS pricing turns into agentic commerce.
5/ @jasonlk already sees the lazy version: raise seat prices AND add large fees for agent and MCP access.
"It works on paper, and it works in the short term. But it does not attract net new customers."
A toll on the agent door teaches agents to route around you.
4/ Where I disagree: "You still charge a SaaS subscription."
Agents don't sit in seats. They make calls on someone's behalf, with a budget. The natural meter is per action, not per head.
3/ He's right about the moat. When an agent is the user, the dashboard stops mattering. What's left is proprietary data, approvals, audit trails and the system of record.
A SaaS product that's mostly UI is in trouble.
2/ His argument: stop building chat UIs that resell lab tokens at thin or negative margins. Expose your data, logic and permissions as tools. Let ChatGPT, Claude or Copilot run the conversation. The customer pays for the tokens.
Best founder story: 75+ investor meetings to raise Robinhood's first $3M. Tim Draper came in with $250K after Vlad and Baiju offered to skip salary until the brokerage license was approved.
Fundraising is a numbers game. Keep going.
For students: while many AI labs say no new grads, he's hiring early-career talent aggressively.
His reasoning: nobody has 5 years of experience building 24/7 agents. In a period of fast change, no experience is an advantage.