AI agents can remove the administrative work that keeps people from starting and running businesses.
2
A group of cooperating agents can accomplish work that one language model cannot handle alone.
3
When agents become widely available, individuals may gain the capabilities of large companies without their money, staff, or networks.
Summary
Flo Crivello argues that AI agents will reduce the friction involved in starting and running a business. Generative AI can create text and images, but agents can take actions such as managing calendars, handling email, preparing for meetings, and using software tools. Crivello describes Lindy as a framework that lets an AI pursue arbitrary goals with arbitrary tools, while acknowledging that it can fail, including when it invented vocabulary words. He then extends the idea from one assistant to a society of agents. Lindy uses specialized agents to find documentation, manage engineering work, write authentication code, run QA, and submit a pull request for an integration. Crivello compares this cooperation with the difference between one person trying to build a toaster and a group producing one cheaply. He predicts that individuals will eventually work with AI marketers, designers, and engineers, giving a person with a laptop the reach of a much larger company.
Agents can remove the friction that makes business administration exhausting
Crivello says the hard part of starting a company is often everything around the product: incorporating, meeting lawyers and bankers, opening accounts, raising money, recruiting, and hiring. Generative AI tools that write copy or make images do not solve that problem. He became interested in agents because they can automate work and actions across a person's life. He frames this as a response to a workplace where people spend time on administrative tasks they dislike, rather than on work that requires their own judgment.
Lindy handles calendars, email, and meeting preparation through remembered preferences
Crivello describes Lindy as an AI employee that first acts as a personal assistant. It can schedule meetings, remember preferences, and apply feedback in future interactions. When Crivello corrected a meeting name from "meeting with Flo" to "Flo Eric," Lindy renamed it and saved the preference. It can also be copied on email, draft replies in the user's voice for different recipients, and prepare for meetings by sending a Zoom link, LinkedIn profiles, and recent email summaries five minutes before each meeting.
A universal agent framework produces features the team did not build one by one
Crivello says Lindy's team did not directly build each assistant feature. They built a universal framework that lets an AI pursue an arbitrary goal with an arbitrary tool. He argues that complex behaviors can emerge from this general setup. The intended computing experience is conversational: people focus on what they uniquely do well while routine work arranges itself. He presents this as a change from spending the day working inside a computer on repetitive tasks.
Crivello deliberately gives an example where Lindy did not work. He asked it to send him an interesting new word every morning. One message defined "p liquidity," a word that did not exist. After checking the other messages, he found that none of the words were real either. The example puts a limit on his broader claims. Agents can automate useful work, but their output still needs checking, especially when they generate information rather than carry out a well-defined action.
A society of agents can divide work that is too large for one model
One assistant is not enough for Crivello's larger vision. He compares a language model asked to design and publish an entire iOS app with one person being asked to build a rocket. Failure by one worker does not show that the whole task is impossible. Humans make complex things by working together, and agents can do the same. Lindy's framework coordinates multiple agents that divide a goal into separate jobs and pass work between specialists.
Lindy uses specialized agents to build its own integrations
Crivello describes a group of Lindy agents working on integrations with tools such as Slack, Twilio, and Google Sheets. A tool-creation agent receives an instruction, another agent finds the OpenAPI specification and web documentation, and a manager divides the task among engineering agents. One engineer handles authentication. QA tests the result and sends it back when it fails. If it works, the system submits a pull request. Crivello says this process was 70 or 80 percent complete when he presented it.
Agents could give individuals capabilities now concentrated inside large companies
Crivello predicts that the strongest marketers, designers, and engineers will eventually be AI systems that work for individuals. He calls this a great equalizer because a person would no longer be limited by time, money, team size, or professional network. His long-term example is a 25-year-old in a San Francisco studio having more business impact than Coca-Cola. The remaining skill, in his view, is knowing how to use this expanding set of capabilities.
"Your Computing experience of the future isn't one where you're working on your computer, it's one where you're having a conversation with your computer."09:08
Who should watch
Founders who spend more time on incorporation, hiring, email, scheduling, and fundraising than on their product will get a concrete picture of how agents might remove that work.
Engineers building agent systems will find an example of multi-agent task division, including documentation search, management, authentication, QA, and pull requests.
People evaluating AI assistants should watch for Crivello's failure example, because it shows why generated information still needs review.
