In this episode of the Get Paid with Manny Medina, Flo Crivello, Founder and CEO of Lindy, shares the full origin story of one of the most ambitious AI products being built today: a proactive executive assistant that manages your inbox, prepares you for every meeting, and takes action on your behalf before you even ask.

Flo traces the journey from a 2022 meeting recorder to a no-code agent builder to the AI chief of staff executives are now replacing their human assistants with. Along the way, he opens up about losing talent through pivots, the dangerous middle ground between product and platform, and the model release that made him realize the guardrails he’d built were now the ceiling holding Lindy back.

Flo began working on AI agents in mid-2022. GPT-4 didn’t exist yet. LangChain didn’t exist. The word ‘agents’ wasn’t part of the industry vocabulary.

The insight came from experimenting with the GPT-3 API while building a meeting recorder. The team realized the model wasn’t just good at generating language; it could take actions.

While the rest of the market rushed to build AI writing tools, Flo was quietly trying to build software that could actually do things.

Early agentic AI was far too unreliable to ship as a freeform system. Hallucination rates were unworkable. Function calling didn’t exist. The team resorted to having models write raw code to hit APIs, a fragile, error-prone approach.

Lindy’s solution was a structured canvas, similar to Zapier, where humans defined every step in a workflow and the agent filled in the blanks. It was rigid. But it worked.

The model proved itself quickly. Within weeks of building on Rails, Lindy was automating complex workflows for a prominent VC firm that had doubted it was possible. Shortly after, a YouTube creator named MattVidPro discovered the product, and the inbound exploded, with most companies spending heavily to manufacture, and arrived for free.

As the agent builder grew, a harder problem emerged. Lindy had drifted into the space between a product and a platform, too opinionated for developers, too technical for end users.

Flo’s analogy is sharp: telling someone you’ve built the world’s easiest way to make their own cheeseburger at home doesn’t land, because people who want a cheeseburger want McDonald's.

The realization forced a real decision: go hard toward developers and compete in the infrastructure space, or go hard toward end users and build something genuinely magical for people who are too busy to configure anything.

Lindy chose the latter. That choice sent them back to the original vision.

Throughout all the pivots, Flo had been running Lindy as his own personal AI assistant, swapping in each new model as it was released and watching the experience slowly improve. When Claude 3.5 arrived, something changed.

End-to-end agents, fully autonomous loops with no human-defined steps, had always been the weakest part of the product. Suddenly, they worked. The structured workflows that had made Lindy reliable were now limiting what the agent could do on its own.

Lindy rebuilt around the vision Flo had been carrying since 2022.

Lindy connects to your email and calendar, learns your context continuously, and acts before you ask. There’s no setup wizard. No flow to configure. It starts delivering value within minutes of connecting your accounts.

During a recent engineering interview, a candidate mentioned a referral. By the time the meeting ended, Lindy had found the LinkedIn profile, drafted a personalized outreach email, and sent Flo a text asking if he wanted it sent.

The breakthrough surface turned out to be iMessage. What started as a feature became the core of the product, a text interface that lets busy executives give and receive information in the same seconds they’re glancing at their phone between back-to-back meetings.

For founders building now, Flo’s advice centers on a constituency most people are still underestimating: agents themselves.

Agents are becoming buyers. They need compute, billing infrastructure, memory, and tooling designed for non-human operators. The founders building that layer, not for developers, not for end users, but specifically for agents, are sitting on a wide-open opportunity.