Lindy CEO Flo Crivello argues that for AI to be a true teammate, it must inhabit shared spaces like Slack and possess a deep, shared context of the team's history. Raw intelligence is less useful without this context, making agents potentially better than humans at onboarding.
Lindy uses an AI agent to continuously manage its own memory. This "meta-memory" agent learns to identify and ignore low-value data sources (like Slack log channels) and prioritize high-signal information, a more dynamic and intelligent approach than static RAG systems.
Lindy handles sensitive data not with rigid access controls, but by allowing users to give natural language instructions in a text file that serves as the agent's "meta-memory prompt." This lets users define complex privacy rules dynamically through simple prompting.
To manage huge context sizes, Lindy uses "recursive context buckets" organized in a self-balancing tree. This data structure allows an AI agent to access information from a context of billions of tokens with just two LLM calls, effectively solving the context window limitation for complex tasks.
Lindy CEO Flo Crivello reveals they subsidize their product but still focus intensely on maintaining an 85% cache rate. He notes a drop from 85% to 65% nearly doubles costs, making caching a critical lever for unit economics, not just a minor optimization.
Lindy CEO Flo Crivello argues that while human-AI collaboration currently yields the best results, this "centaur phase" is fleeting. Citing AI's history in games like chess, he predicts that humans will eventually detract from, rather than enhance, highly optimized AI systems.
Lindy's CEO advises against creating multi-agent systems that mirror human job roles (e.g., designer, PM). This is a flawed anthropomorphism, as AIs lack human constraints on time and context. A single, powerful agent is often more efficient than a team of specialized agents.
When multiple AI agents need to modify a shared knowledge base, Lindy avoids complex database transaction logic by using a Git-backed file system. This allows agents to handle collaboration issues like merge conflicts using a battle-tested paradigm from human software development.
Flo Crivello describes frontier AI as a superintelligence that can write 50,000 lines of code but then makes absurdly simple errors. The key challenge in human-AI hybrids is designing systems where the AI knows when it's about to be dumb and can escalate to a human.
Lindy dramatically increases agent reliability with a "validator" system. Before an action is taken, a second LLM call acts as a judge, checking the proposed action against an extensive prompt or checklist. Even a simple "Are you sure?" prompt provides a significant reliability bump.
Despite founder Flo Crivello's public calls to ban Chinese models, his company Lindy runs its core AI Teammate product on DeepSeek by default. The model is roughly at the Sonnet 4.6 level but significantly cheaper, making the economics undeniable for a high-volume agentic product.
Lindy has a homegrown automated system that runs thousands of evals to re-optimize its system prompts for each new major model release. This costs ~$10,000 per model, a necessary expense as prompts that work for one model family often fail on another, highlighting a hidden operational cost.
Flo Crivello advocates banning Chinese AI models in the U.S., citing unfair competition from distilled models, the risk of CCP propaganda and economic influence, and the need to protect America's domestic AI industry. He frames it as a matter of fairness and national security.