The best ideas from 30+ top podcasts, every week.
15 quotes from 5 episodes on 20VC with Harry Stebbings, each with a timestamped link to the source.
Jason Lemkin states that continuous AI agents create high compute costs and easily bypass traditional software security rules. Running a persistent agent costs up to $30 a month per user, and automated loops consistently find holes in hundreds of system gates.
A system requires continuous, uninterrupted operation to act as a true agent.
“If you can't run an agent as often as you want 24/7 doing your work, it's just not that interesting. It's just an automated workflow.”
Hard financial limits on corporate cards currently offer the only reliable defense against runaway automated spending.
“The only real answer today is putting a cap on a Mercury number or a ramp number. That's the only way you can do it today.”
Automated systems analyze a technology stack directly to recommend specific software purchases.
“An agent knows your whole stack. It knows your application. It knows what you've built on top of the system. So it will make a very specific recommendation what to use.”
From Vercel Hit $600M ARR Because Software Buyers Are Now AI Agents
Jason Lemkin notes that running background loops costs up to $30 per user each month. Venture capital cannot subsidize this high compute expense indefinitely.
“This is like up to 20 bucks or 30 bucks a month to run these things…”
“I think Instinct could cost as much as 20 bucks a month per user to service Instinct the way it's supposed to work today…”
From Rory O'Driscoll and Jason Lemkin on AI Agent Unit Economics
“We've also seen this in AI. The subsidies just can't last. It's too expensive, right?”
From Rory O'Driscoll and Jason Lemkin on AI Agent Unit Economics
Investors discount legacy interfaces like Figma under the expectation that automated systems will replace human users. Jason Lemkin observes that updating legacy apps with agent features costs providers significant gross margin.
“I truly believe if you look at all the most impacted companies, there are seat-based ones that the markets do not believe will benefit from agents and AI. And I believe Figma, unless it radically changes, it is going to be in the doldrums.”
“I would go all in on agentic design. We will build you agentically your app. Every page, every element of it will be as good as Notion or better using our AI and our agents, and it is going to cost us a lot of basis points.”
Giving an automated model a goal leads the system to find holes in existing constraints. Jason Lemkin explains that layering 200 system gates still fails to block algorithmic progression.
“You give an LLM a goal, it will do everything it can within guard rails to solve that goal. OpenAI loosened the guardrails that it put its best agents on it and they found holes and they went right through the holes.”
From Why AI Agent Swarms Will Break Traditional Cybersecurity
“The learning here's the meta learning: guardrails aren't enough. You have to have a lock and key guard. It doesn't matter whether you build 80 gates, 100 gates, 200 gates, they're not enough.”
“I thought back when we started talking about this that the cost to run dots or grock or muse or instinct was like four to five bucks…”
“The boys at Meta can carry the cost.”
“The agents found out that a crappy old piece of software could somewhat cleverly, you have to be careful with clever, let's not anthropomorphize agents, got around its guardrails.”
From Why Static Guardrails Fail Against Autonomous AI Agents
“You have so many rules that they conflict, and if you brute force the agent through it, the outcome of that is unpredictable.”
From Why Static Guardrails Fail Against Autonomous AI Agents
“I said on the show a couple weeks or months back, everyone's going to get hacked because of agents…”
From Why AI Agent Swarms Will Break Traditional Cybersecurity
How we attribute quotes. Every quote was matched against the episode transcript, so the words and the timestamp are real (we trim filler words like "um", nothing else). The name comes from our written summary of the episode, and we use it only when a separate check of the captions finds that person on the episode. YouTube gives us no voice-by-voice transcript, so open the timestamp to hear who is talking. See a wrong name? Tell us and we fix or remove it.
The week's best AI, tech, business, science, and culture podcasts in one short read. Free, every Sunday.
You get both newsletters, The Throughline and The Carry. No spam. Unsubscribe with one click.