Lindy.ai

Lindy.ai builds AI assistants and a no-code platform for automating business work. Founded by Florent “Flo” Crivello, its products cover email, scheduling, meetings, customer support, and cooperating agents. A consequential change was its 2024 rebuild, Lindy 2.0, which replaced a giant prompt and tools with explicit workflow steps. Later offerings added cloud computers for agents, web-app generation, and phone assistance.[1][6][7][8]

Before the workflow rebuild, Lindy began as a general assistant. Crivello recalled starting the project at the end of 2022. Its March 2023 introduction presented calendar management, email triage and drafting, meeting recording, and recurring workflows triggered by emails or calendar invitations. The company described guardrails that considered uncertainty and the consequences of an action when deciding whether to proceed or request confirmation.[1][4]

The November 5, 2023 announcement broadened this into a platform for teams of agents instructed in plain English. Agents could wake on emails, tickets, or webhooks and work together. One company-described example assigned separate agents to find API documentation, distribute integration work, write code, and perform QA. Feedback was stored in inspectable plain English. The announcement promised availability soon; it did not establish general availability that day.[5]

At Latent Space’s February 2024 anniversary event, published in a March 9 episode, Lindy was introduced as recently opened to the general public. Crivello reported more than 3,000 integrations and demonstrated a support agent meant to pass ticket details to a second agent that logged them in a spreadsheet. The support agent answered the refund request but failed to delegate; Crivello acknowledged the failure during the demo.[2]

In the November 15, 2024 interview, Crivello described roughly six months spent rebuilding the product as Lindy 2.0, also called “Lindy on Rails,” with users receiving it over approximately the preceding two months. He judged the new version more reliable and easier to configure: users selected steps rather than trying to express an entire automation in one large text field.[1]

The Zendesk example explains the architectural change. Previously, a prompt asked the model to consult a knowledge base before answering a ticket, leaving the model to decide whether to invoke it. The rebuilt workflow placed knowledge-base consultation as a required step between the ticket trigger and the agent’s response. This made consultation part of the execution sequence; it did not guarantee that every answer would be correct. Crivello described the approach as “using as much AI as necessary and as little as possible.”[1]

Individual action fields still used prompts. In Crivello’s reservation-email demonstration, a workflow filtered booking confirmations and appended spreadsheet rows, extracting the reservation’s start date, amount, and traveler count. Users could choose a model for each node. Structured outputs were used where possible, while some message-generation steps produced plain text.[1]

Workflows also retained context for later interaction. Crivello’s meeting recorder sent summaries and coaching notes, then restored the meeting and workflow context when he replied to an email or asked a question in a Slack thread. Agents could save memories between executions and call other Lindys. Crivello manually pruned his own agents’ memories because excessive memory could confuse them, though he suspected this was less necessary with Claude 3.5 Sonnet.[1]

Lindy built internal infrastructure to turn conversation histories into evaluations in one click. Good conversations became examples of expected behavior; negative examples could be edited after generation. By November 2024, Crivello considered the homebuilt evaluation tool burdensome to maintain and was contemplating a move to Braintrust, which Lindy had not yet adopted.[1]

Crivello recounted a customer-support agent sending a rickroll link when asked for video tutorials that did not then exist. The team sent a correction before the customer opened the first email. A log query found, in his recollection, three or four earlier instances. He said the team addressed the behavior with an added system-prompt instruction.[1]

Users could require confirmation before an individual workflow step and approve, deny, or edit an action before execution. Lindy was developing a further mechanism to embed user edits in a vector database and retrieve them as examples for future tasks. This was contextual feedback retrieval under development, rather than evidence of training model weights.[1]

Crivello described an append-only context ledger: nodes added information that subsequent nodes inherited, allowing Anthropic-managed prompt caching to work with the growing prompt. He credited Claude 3.5 Sonnet with making model capability less of a bottleneck and viewed integrations and product engineering as the more pressing constraints. His roadmap favored API calls wherever available, citing their performance and reliability advantages over computer use.[1]

On July 30, 2025, Lindy announced Lindy 3.0 as available, combining three components: Agent Builder created agents from prompts; Autopilot gave agents cloud computers they could see and operate; and Team Accounts provided centralized sharing, deployment, management, and monitoring. Autopilot extended the platform beyond API integrations. The company reported more than 6,000 integrations, a count of integrations rather than customers or users.[6]

Lindy Build, announced as available on August 26, 2025, expanded into generating web apps from prompts, with databases, integrations, frontend, and backend configuration. Its company-described testing mechanism clicked through generated apps and exercised forms, interfaces, and features while finding and fixing bugs. Build integrated with Lindy Agents so app creation could be paired with business automations.[7]

On October 2, 2025, Lindy launched Gaia, a phone agent powered by Deepgram Flux. Its described uses included answering inbound support calls using a knowledge base, screening calls, scheduling appointments, and escalating to team members. Gaia was a voice product.[8]

The February 6, 2026 Lindy Assistant announcement packaged email triage, reply drafting, calendar management, meeting preparation, notes, and follow-ups into a live assistant offering. Users could interact through iMessage, email, Slack, or a browser. The described inbox workflow left review and sending to the user.[9]

Crivello argued that a horizontal agent platform could share engineering work across use cases and let agents cooperate under one roof, while still leaving room for vertical products. By November 2024, Lindy had hired general managers to own individual verticals’ go-to-market work and template customization. He said the arrangement drew on his earlier experience with geographically organized general managers at Uber.[1]

Crivello also allowed the team to adopt Linear despite preferring Notion’s unified view of roadmaps, tasks, and postmortems. He explained that the team needed to choose its own tools, accepting that the change made the company somewhat less legible to him.[1]

Lindy’s work arrangements marked a reversal from Crivello’s earlier business, Teamflow, which built a virtual office for remote teams. In the November 2024 interview, he recalled strongly advocating remote work before concluding that it made collaboration, culture, and alignment on the product harder. He said Lindy had stopped working remotely. This was his assessment of the company’s needs and his earlier experience.[1]

Lindy operates under Crivello Corp, doing business as Lindy. Its terms of service, displaying a March 14, 2025 update, identify the company as registered in Oregon and as the operator of lindy.ai.[3]