CHAPTERS
Why AI agents can be “100x cheaper and better” than humans
Aakash introduces Lindy and CEO Flo Crivello, framing AI agents as the next step in automation where software can outperform people in specific roles. Flo argues this shouldn’t be controversial—like calculators vs. humans—because the range of tasks computers dominate keeps expanding.
Agent #1: Meeting recording as a “second brain” (and why Lindy differs)
Flo demos his meeting recording agent: it records, summarizes, and lets him query past meetings for details (e.g., hackathon sponsorship amount). The differentiation vs. standalone note apps is customization—Lindy can route summaries into the correct Slack project channel automatically.
Reducing LLM unreliability: human-in-the-loop confirmations
Aakash challenges reliability and “lossiness” of LLM-driven automation. Flo argues models have improved quickly and aren’t worse than humans for many tasks, and shows a one-click human approval step to prevent brand-damaging actions.
Choosing models by speed vs. intelligence (and cost tradeoffs)
Flo explains model selection as a pragmatic optimization: use fast models for low-latency tasks and the smartest models for high-stakes decisions. He shares examples including phone agents and a betting agent that spends dollars in tokens to make better calls.
Building an agent from scratch with Agent Builder: Email triage in minutes
Flo uses Lindy’s natural-language Agent Builder to create an email triage agent for Gmail that labels messages (urgent/FYI/archive/investor). The discussion highlights “context engineering” via editable external docs (e.g., a Google Doc policy the agent consults).
Agent #2: CRM manager that proactively surfaces relationships and follow-ups
Flo describes a CRM agent that remembers people he meets, answers queries like “who are the salespeople I know,” and proactively suggests contacts when travel is detected. He also shows the builder can recreate complex agents by asking clarifying questions.
6,000 integrations plus “Computer Use” to bypass missing APIs
Aakash asks about integrating CRMs and other tools; Flo emphasizes breadth of integrations and introduces “computer use” as a fallback when APIs don’t exist. He demos an agent that blocks Twitter/X mention spammers by navigating the UI directly.
Cost reality of Computer Use and token/context mechanics
Flo and Aakash discuss whether UI automation is expensive; Flo says it’s often cheaper than expected due to different context handling. He notes costs are largely linear with tokens and previewed guardrails to cap spend per task.
Agents for business operations: Support and Recruiting (plus Agent Swarms)
Flo highlights support automation as the highest-leverage business use case because it scales elastically (Black Friday spikes) with near-instant responses. He then demos recruiting agents that find candidates and use “Agent Swarm” to execute outreach reliably in parallel.
Hardening workflows: deliverability guardrails and measuring agent performance
Flo explains where agents can ‘mess up’ in real operations, like triggering Gmail spam blocks during mass outreach. He shares patterns for tracking campaigns in spreadsheets and for monitoring outcomes via logs and weekly digests (including a personal weekly meeting recap email).
PM-specific agent stack: alignment, VoC digest, virtual user, and “mini-PM” decision agent
Asked what he’d build as a PM, Flo prioritizes meeting dissemination for cross-functional alignment and a daily voice-of-customer digest from support tickets. He also describes a “virtual user” agent built from internal knowledge bases and a shareable ‘mini-PM’ agent trained on a PM’s principles and prior decisions.
Safety, permissions, and the shift from doer to orchestrator
They cover how to prevent agents from harming brand or systems: restrict tool access and add confirmations for risky actions. Flo frames the broader work shift as moving from execution to managing an “empire” of agents—similar to getting promoted to manager, but without most of the ‘soft’ people-management burden.
Platforms: ChatGPT/Claude vs Lindy; Lindy vs Zapier/Make/n8n
Flo compares consumer single-agent tools (ChatGPT/Claude) to workplace orchestration platforms (Lindy) with richer controls. He argues Lindy is more AI-native than workflow tools like Zapier and easier than developer-oriented automation stacks, emphasizing capability plus ease of use.
Company-building lessons: hiring story, founding/pivot journey, and AI transformation playbook
The conversation widens to startup operations: Flo recounts hiring/firing a high-profile engineer who was moonlighting, and shares hiring heuristics (trust gut, avoid job-hopper patterns, do backchannels). He then tells Lindy’s origin from a prior startup pivot, argues strategy is emergent, and outlines how companies should approach AI transformation (COE/‘AI czar’ role done full-time).
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome