Florent Crivello

Florent Crivello, also known as Flo Crivello, founded Lindy, a platform for building AI agents without code. A consequential decision in that work was his team's rebuild of Lindy around explicit workflows rather than a broad prompt and a collection of tools. In his November 2024 interview, he argued that constraining agents made them both more reliable and easier for users to configure. Lindy's July 2025 announcement identified him as its founder and CEO.[2][5]

Crivello described Lindy 2.0 as a substantial rewrite. The earlier interface asked users to explain an automation in a large text field; the replacement let them select triggers and actions step by step. His Zendesk example explains the technical distinction: a prompt could ask an agent to consult a knowledge base before answering a ticket, but leave the model to decide whether to invoke retrieval. An explicit workflow placed knowledge-base consultation between ticket receipt and the agent's response, making it a required step.[2]

His principle was to use “as much AI as necessary and as little as possible.” He endorsed Swyx's characterization of putting the “Shoggoth in a box,” while describing incremental permissions as another constraint: an email assistant requested authorization to read email, then separately requested permission to write drafts. This could require repeated authorization, but followed his preference for requesting only the access needed at each stage.[2]

Crivello's own automations illustrated where language models fit within those workflows. He demonstrated extracting dates, payment amounts and traveler counts from reservation-confirmation emails into spreadsheet rows, avoiding a custom HTML parser. His meeting assistant sent summaries and coaching notes, then restored the meeting's context when he replied with a follow-up question. In one example, he asked why it had found his sales explanation too technical; it pointed to terms such as deterministic behavior, context windows and agent state.[2]

He also described the work needed to keep agents useful over time. He personally pruned his assistant's memories because too many could confuse it, while acknowledging that pruning might be less necessary with newer models. His team built infrastructure to turn conversation histories into evaluations in one click, editing negative examples to show the desired behavior. By November 2024, he regretted maintaining a custom evaluation tool and was considering Braintrust; the migration had not yet happened.[2]

Two public examples exposed the limits behind the agent vision. At Latent Space's February 2024 anniversary demonstration, published in March, Crivello configured a support agent to pass ticket details to a reporting agent that would log them in a spreadsheet. The support agent answered, but he acknowledged: “She did not delegate.” In November, he recounted a support agent sending a rickroll link when asked for video tutorials that did not yet exist. His team corrected the response, searched for other occurrences and added an instruction against rickrolling.[3][2]

For Lindy's workload, Crivello identified 3.5 Sonnet as the point when the model ceased to be the main bottleneck. He described an append-only context ledger: each workflow node added information, and subsequent nodes inherited the accumulated context. That growing prefix worked well with prompt caching managed by Anthropic. His assessment shifted attention toward integrations and product capabilities, rather than treating model capability as the explanation for every limitation.[2]

One capability under development was learning from user corrections through retrieval. Users could already require confirmation before a workflow step, then approve, reject or edit a proposed action. Crivello said the team was working on saving those edits in a vector database and retrieving them as examples for future tasks. His description concerned supplying examples in the model's context, rather than training its weights.[2]

In the same interview, he favored APIs for tasks that APIs could perform, arguing that their performance and reliability advantages over computer-interface use remained consequential. His later Lindy 3.0 announcement, dated July 30, 2025, introduced Autopilot, giving agents cloud computers they could see and operate to reach beyond API integrations. He also presented prompt-based Agent Builder as a way to simplify agent creation, extending his earlier focus on making agents easier for users to configure.[2][5]

Crivello's broader thesis favored a horizontal agent platform. He compared agents with search: shared technical problems and the expense of solving them could justify spreading research costs across many uses. Keeping agents on one platform also let them work together and reduced the number of systems users had to learn. He explicitly rejected an either-or outcome, predicting both specialized agents within software categories and large horizontal platforms.[2]

Before Lindy, Crivello worked at Uber and founded Teamflow, a virtual-office business for remote teams. Those earlier experiences supplied concrete precedents for his later management choices and his reassessment of remote collaboration.[6][2]

In an April 2023 interview, Crivello recalled spending almost five years at Uber, first as an engineer and then as a product leader. He said founding a startup had already been his goal: he wanted a green card, savings, skills and a network before starting. He linked his move into product management to preparing for entrepreneurship, though in hindsight he felt he had waited too long. He credited Uber with teaching him the value of giving people substantial ownership and holding them accountable.[6]

Lindy began with an ambition to automate knowledge work, but Crivello's team chose executive assistance as its initial market. He said they considered law and finance, where they found demand, but judged those to be substantial categories in their own right rather than the starting point they wanted. Calendars, email and meeting notes offered a more tractable entry. His introductory Lindy article, dated March 21, 2023, put reliability ahead of proactivity and described weighing uncertainty and the consequences of an incorrect action before requesting confirmation.[6][4]

In April 2023, Crivello described a different technical architecture from the later workflow system: Lindy generated task-specific code and built integrations from API documentation. His team evaluated and fine-tuned base models, using models specialized for dialogue, factual questions, browsing and code writing. He framed the generated code as a temporary application for the user's immediate task—for example, displaying calendar events on a map.[6]

By November 2024, Crivello said Lindy's use of general managers for individual verticals drew on Uber's geographic GM structure. He continued to spend substantial time on product work. He also recounted letting his team adopt Linear despite preferring Notion's consolidated view of roadmaps, tasks and postmortems: the team needed to self-organize and choose its tools, even if that made the company less legible to him as founder.[2]

His reversal on remote work was especially pointed because Teamflow had been built to support it. In November 2024, he said remote hiring, lower rent and avoiding commutes had clear benefits, but that he had found collaboration, culture and alignment on the product's underlying mental models harder. He distinguished optimizing known work for cost from deciding what to build, which he considered a creative endeavor better served by working together in person. This was his explanation for changing his own position, rather than a demonstrated universal limit on remote teams.[2]

Crivello also argued that career-focused technology founders, particularly in AI, should be in San Francisco because of its network effects. On AI safety, he said he had publicly supported SB1047 and described concern about risks at the model layer alongside enthusiasm for AI's potential upside. He defended speaking openly about unpopular views as a response to groupthink, and presented participation in building AI as compatible with taking its risks seriously.[2]