Agents @ Work: Lindy.ai

Episode Summary

Executive Summary: The episode centers on Lindy.ai founder Florent Crivello’s view that AI agents work best when constrained by deterministic workflows, not free-form prompts. He demos Lindy 2.0’s “on rails” no-code agent builder, discusses permissions, memory, evals, and use cases like meeting notes, inbox processing, support, and scheduling, and argues that horizontal agent platforms will matter broadly as models improve. The conversation also covers remote work, SF vs Europe, contrarian public speech, and AI safety.

Main Topics: Lindy 2.0 and the “on rails” agent paradigm (Priority: 5/5): Crivello explains that Lindy was rebuilt from a giant prompt into a workflow system with explicit triggers, actions, and guardrails. The goal is higher reliability and easier setup for non-engineers. No-code AI agents for business automation (Priority: 5/5): Lindy is positioned as a no-code platform for building AI agents that automate workflows like email triage, meeting prep, support tickets, scheduling, and reporting. Permissions, memory, and reliability tradeoffs (Priority: 4/5): The discussion focuses on least-privilege OAuth, incremental permissions, memory pruning, and why too much autonomy or memory can reduce agent quality. Product demos and real-world use cases (Priority: 4/5): Crivello walks through concrete Lindy examples: reservation logging, meeting coaching, Slack dissemination, inbox processing, blood pressure logging, podcast summaries, and PR review. Horizontal vs vertical AI products (Priority: 4/5): He argues that many agent capabilities will converge into a few horizontal platforms, even though vertical tools will emerge first because they are easier to build. Company building, remote work, and organizational design (Priority: 3/5): The conversation broadens to why Lindy moved back to in-person work, how GMs are used to manage verticals, and how legibility vs performance affects company structure. Contrarian views, SF, Europe, and AI safety (Priority: 3/5): Crivello discusses speaking openly on controversial topics, his pro-SF and anti-Europe stance for founders, and his belief that AI upside is huge but model-layer safety risks remain real.

Key Arguments: AI agents become more reliable when the workflow is explicitly structured; prompts alone are too ambiguous and brittle. Users often do not know how to express tasks in text, so GUI-based step-by-step configuration is better than pure prompt interfaces. Least-privilege, incremental permissions are preferable to broad access, even if setup becomes slightly more cumbersome. Memory should be curated and pruned; more memory is not always better because agents can become confused by irrelevant facts. A horizontal agent platform can win because agent problems share more in common with each other than with any one vertical domain. The model is no longer the main bottleneck for many agent products; product, integrations, and workflow design matter more now. Remote work is better for cost efficiency, but in-person work is better for creativity, alignment, and product building. Speaking openly and resisting groupthink is valuable because public silence creates false consensus and weakens judgment. AI safety concerns are real, but they do not negate building useful products; the upside is large enough to justify participation with caution.

Data Points: Lindy 2.0 rebuild timeline: ~6 months - Crivello says the product was rebuilt from scratch over the last six months before the new version was rolled out. User rollout of Lindy on Rails: ~2 months - He says users have been using the new on-rails version for roughly two months. System prompt size: >4,000 tokens - He notes Lindy’s system prompt is now larger than the 4,000-token context windows that were common when the company started. Model bottleneck threshold: Claude 3.5 Sonnet - He says 3.5 Sonnet was the point where models stopped being the main bottleneck for Lindy. AIME score comparison: O1 ~83 vs 3.5 Sonnet ~14 - Used to illustrate that O1 still outperforms 3.5 Sonnet on some reasoning benchmarks. Google Workspace account cost: ~$10/month - He says some users provision separate Google accounts for AI agents at roughly this monthly cost. Meeting scheduling action size: ~1,000 lines of code - He describes the find-available-times connector as surprisingly complex and large. Guidelines length: ~40 pages - He says Lindy’s PR-review guidelines have grown to about 40 pages. Customer support recall incidents: 3-4 instances - He says they found a few cases where Lindy accidentally rickrolled customers before fixing it. AI safety personal estimate: ~10% p(doom) - Crivello gives a rough subjective estimate of existential downside risk from AI.

Pivotal Quotes: "put Shoggoth in a box and make it a very small, like the minimal viable box" — Alessio: A metaphor for constraining the model with deterministic workflow rails instead of relying on free-form prompting. "I think text is awesome. And I've actually come around. I actually sort of agree now that text is really not great." — Florent Crivello: He explains why GUI-based agent setup is often better than pure text prompts for real users. "If you are in tech, especially in AI, but if you're in tech and you're not in San Francisco, you either lack judgment or you lack ambition." — Florent Crivello: His strongest statement on founder geography and the importance of SF for ambitious AI builders.

Implications: The episode suggests agent platforms will win by combining model power with strong product constraints, permissions, and UX. For builders, the near-term edge is workflow design and integrations, not raw prompting. For founders, SF, in-person collaboration, and open debate remain central themes.