CHAPTERS
- 0:00 – 3:16
Why AI agents can be 100× cheaper and better than humans (and why adoption is still low)
Aakash introduces Lindy and CEO Flo Crivello, framing AI agents as a near-term way to replace certain team roles. Flo argues this is a natural continuation of automation, yet most working PMs still haven’t built an agent—creating an opportunity for early adopters.
- •AI agents as replacements for specific roles, not just assistants
- •Flo’s claim: 100× effectiveness/cost for certain tasks
- •Tech bubble vs real-world adoption gap (only ~8% in PM community)
- •Positioning 2025 as a major monetization window for agent builders
- 3:16 – 6:57
Agent #1: Meeting recording as a ‘second brain’ with automated follow-ups
Flo demos his meeting-recording agent that stores searchable meeting history and answers questions later (e.g., commitments made in a call). The differentiator versus note-taking apps is deep customization—like auto-posting summaries to the right Slack project channel.
- •Meeting recorder as durable memory: query any past meeting
- •Auto-generating and distributing action items/summaries
- •Conditional routing: detect project meetings and post to matching Slack channels
- •Customization is the key advantage over simple note-takers
- 6:57 – 8:54
Reliability & brand safety: hallucinations, human-in-the-loop approvals, and learning from edits
Aakash presses on LLM unreliability and misrouting risk. Flo argues models have improved rapidly and can be governed using human-in-the-loop confirmation steps that let users approve sensitive actions and (eventually) teach the agent from corrections.
- •LLM ‘hallucination’ reputation is partly legacy from early GPT-3.5 era
- •Agents are non-deterministic but comparable to humans in error rate for many tasks
- •One-click human approval before executing actions (e.g., Slack posting)
- •Future: agents learn from user edits to reduce repeat mistakes
- 8:54 – 10:32
Choosing the right model: speed vs intelligence vs cost (plus token ‘burn’ for high-stakes tasks)
Flo explains model selection heuristics: default is fine most of the time, but use fast models for low-latency needs and smartest models for complex reasoning. He shares an example of an agent that researches and places bets, intentionally spending more tokens to improve decision quality.
- •Most users should stick with the platform default model
- •Fast models for phone/real-time scenarios; smarter models for deep reasoning
- •High-stakes tasks justify higher token spend
- •Two main axes: latency and ‘IQ’ (capability)
- 10:32 – 13:26
Agent builder walkthrough: creating an email triage agent in plain English (and ‘agents building agents’)
Flo builds an email labeling agent from scratch using natural language prompts. The platform turns the prompt into a trigger + analysis step + action, illustrating how agent-builder tooling reduces the need for manual flow construction.
- •Start-from-scratch agent creation via plain-English instructions
- •Email triage labels (urgent/FYI/archive/investor) as a simple starter
- •Flow editor enables personalization/context engineering
- •‘Agent prompting an agent’: builder generates the underlying logic
- 13:26 – 17:25
Agent #2–#3: CRM manager agent + 6,000 integrations (and using docs as living instructions)
Flo describes a CRM agent that remembers contacts, proactively surfaces who to meet when traveling, and prompts weekly cleanup by cross-checking calendar vs CRM. He also explains how instructions can live in a Google Doc for continuous updates and highlights Lindy’s broad integration coverage.
- •CRM agent: add/retrieve contacts, proactive relationship reminders
- •Weekly reconciliation: detect people met who aren’t in CRM and prompt additions
- •Use a Google Doc as the updatable policy/instructions source
- •Integration breadth (6,000+) reduces setup friction
- 17:25 – 20:42
Computer Use: bypassing missing APIs (Twitter spam blocking) and surprising cost dynamics
Flo introduces ‘computer use’—agents operating a cloud browser/desktop to do anything a human can, even without APIs. He demos a Twitter-mentions spam blocker and explains why computer-use tasks can be cheaper due to constrained context (e.g., only recent screenshots).
- •Computer use enables automation where APIs don’t exist (Twitter/LinkedIn patterns)
- •Example: wake every 3 hours, scan mentions, block spam accounts
- •Cost mechanics: fewer tokens via screenshot-based context vs full API context
- •Practical takeaway: ‘no integration’ stops being a bottleneck
- 20:42 – 24:46
Business agents that replace real headcount: support automation, elastic scaling, and recruiting at scale
Flo argues support is the most obvious ROI: instant 24/7 responses and elastic scaling (e.g., Black Friday spikes) like ‘AWS for labor.’ He then demos recruiting: sourcing candidates and running parallelized outreach workflows using Agent Swarm to avoid bulk-task degradation.
- •Support agents: 30-second responses, 24/7/365, elastic to demand spikes
- •‘AWS for labor’ analogy: pay only for usage, scale instantly
- •Recruiting agent: find candidates (e.g., engineers from Zapier) with real links
- •Agent Swarm: parallel outreach to prevent slowdown/errors in long bulk runs
- 24:46 – 26:54
Operational guardrails & evaluation: avoiding Gmail bans, activity logging, and weekly digests
Aakash asks how to prevent failures and improve over time. Flo shares practical safeguards like tracking outreach campaigns in spreadsheets to avoid Gmail lockouts, and using logs/digests (including a weekly meeting digest email) to monitor performance and iterate.
- •Risk management example: outreach pacing to avoid Gmail deactivation/spam
- •Campaign tracking in spreadsheets and queueing outreach when limits are hit
- •Evaluation via task review, logs, and automated reporting
- •Weekly digest from meeting notes as executive reflection tool
- 26:54 – 31:06
More ‘replacement-grade’ agents: sales prospecting, LinkedIn outreach, and cross-post marketing automation
Flo outlines sales as a massive agent category—especially now that computer use enables DMs, forms, and scheduling without APIs. He also shares a marketing agent that selects appropriate tweets to cross-post to LinkedIn, with human approval to maintain platform-appropriate tone.
- •Sales outreach via contact forms, LinkedIn DMs, Calendly booking
- •Computer use unlocks channels traditionally blocked by missing APIs
- •Marketing agent filters tweets for LinkedIn-appropriate cross-posting
- •Human review step ensures brand/tone safety before posting
- 31:06 – 35:52
PM-specific agent stack: voice-of-customer digests, virtual users, and a ‘mini-PM’ decision agent
Asked what he’d build as a PM, Flo prioritizes meeting capture and distribution, plus daily voice-of-customer summaries from support tickets. He introduces a ‘virtual user’ knowledge base and an internal ‘mini-PM’ agent trained on a PM’s principles that answers engineers and escalates uncertain cases to the real PM.
- •Meeting notes routed to stakeholders reduce misalignment and dropped balls
- •Daily VOC digest: top issues/confusions posted to Slack
- •Virtual user agent: searchable internal expert on users/features/competitors
- •Mini-PM agent: codifies decision principles, escalates edge cases, self-improves
- 35:52 – 40:14
Frameworks for when to build agents: repetitive work, dislike-to-do tasks, and safety via permissions
Aakash and Flo generalize when agents make sense: repetitive/manual work, tasks you could train a VA to do, or anything you don’t enjoy. They cover safety basics—agents can only act within granted permissions—and the mindset shift to ‘orchestrator’ with human approvals for critical steps.
- •Heuristic: never do the same thing twice—automate repeat patterns
- •Automate what you dislike and what scales across the company
- •Permission scoping limits damage; approvals for risky actions
- •Role shift: from doer to orchestrator managing an ‘empire’ of agents
- 40:14 – 44:01
Agent platforms vs chatbots: orchestration, permissions, performance, and cost management
Flo contrasts ChatGPT/Claude as single-agent tools with Lindy-style platforms for orchestrating many collaborating agents. He defines three management dimensions—performance, permissions, and cost—and previews cost guardrails like alerting when tasks exceed a threshold.
- •Chatbots = single-agent personal tools; platforms = multi-agent work systems
- •Orchestration pillars: performance management, permission management, cost management
- •Cost scales with tokens/context; guardrails can request approval above thresholds
- •Granular controls differentiate agent platforms from general LLM interfaces
- 44:01 – 52:59
Competition & company building: Lindy vs Make/Zapier, lean teams with agents, and the Soham hiring story
Flo positions Lindy as more AI-native and easier than developer-oriented automation tools, and distinct from workflow-first tools like Zapier. He shares how Lindy runs lean with internal agents (code review, postmortems) and recounts hiring then quickly firing Soham Parekh, extracting hiring-process lessons.
- •Differentiation: AI employees vs workflow automations; ease-of-use emphasis
- •Lean execution: internal agents for code review and postmortem learning loops
- •Hiring cautionary tale: Soham Parekh ‘multi-job’ scam and rapid offboarding
- •Advice: trust your gut, avoid job-hopper patterns, do backchannel references
- 52:59 – 1:07:41
Founding Lindy early, pivot lessons, and why ‘strategy is emergent’
Flo explains his long-standing AGI interest and how GPT-3.5’s API plus a struggling prior startup led to the pivot. He emphasizes action over planning: start building, let strategy emerge, and pivot faster than feels comfortable to overcome status quo bias.
- •From TeamFlow to Lindy: meeting recorder → CRM updates → generalized agent platform
- •Key unlock: LLMs can ‘do’ via API calls, not just generate text
- •Advice: grand plans are wrong; action produces information
- •Pivot heuristics: follow what pulls you, counter status quo bias, move sooner
- 1:07:41 – 1:16:42
AI agent agencies & enterprise transformation: AI ‘czars’, centers of excellence, and the road to autonomous companies
Flo validates ‘AI agent agency’ businesses as lucrative, comparing today to past digital transformation waves. He suggests successful adoption often requires a dedicated full-time internal leader, and discusses remaining gaps—model coherence over long horizons and effective use of large context/memory.
- •High demand for AI transformation services; early-adopter arbitrage window
- •Pattern: appoint a full-time internal AI lead (not a halftime role)
- •Roadmap gaps: long-horizon coherence, memory, and effective context utilization
- •Workarounds today: context-clearing steps and extracted summaries; optimism on near-term breakthroughs
