Why AI Agents Need Local Computers | Sarah Sachs artwork

Why AI Agents Need Local Computers | Sarah Sachs

MTS

October 16, 2026

Sarah Sachs, Head of AI at Notion, discusses Notion's architecture strategy for AI agents using Vercel Workers and Sandboxes.

0:00

Sarah Sachs

We have a strong perspective at Notion that a lot of things should be served on CPU that are served on GPU. There's a lot of benefit and determinism and executed code in doing important tasks. And Vercel has been a key partner in us being able to bring that vision to our customers in a way that's cost effective and higher quality.

0:18

SPEAKER_2

I love how you worded that. Hello everyone, and welcome back to MTS. We are here live at the Vercel Ship Conference, and I am here joined by Sarah Sachs, who is the head of AI at Notion. Sarah, welcome to MTS.

0:32

Sarah Sachs

Thank you so much. Long time listener, first time attendant. First time here.

0:37

SPEAKER_2

That's how we love them. I will say, I know that you're going on stage later today. You're speaking about all the things that are happening with Notion and Vercel. Can you give us a sneak preview about what you plan to talk about?

0:48

Sarah Sachs

Yeah, I mean, Notion has had a close partnership with Vercel for a long time. We've been customers of each other for a long time. And so we have gone through this agent transformation together, which is really interesting to have that customer-to-customer relationship and user-to-user relationship. So, you know, we have a strong perspective at Notion that a lot of things should be served on CPU that are served on GPU.

There's a lot of benefit and determinism and executed code in doing important tasks. And Vercel has been a key partner in us being able to bring that vision to our customers in a way that's cost effective and higher quality.

1:22

SPEAKER_2

Wait, I love how you worded that. More things should be done on CPU that right now people think should be done on GPU. Can you just make that more concrete?

1:30

Sarah Sachs

Yeah. I mean, so for instance, like having an agent figure out how to write a specific SQL query from that SQL query, then maybe use the Notion CLI to update a dashboard.

1:41

SPEAKER_2

Yeah.

1:43

Sarah Sachs

Agents don't need LLMs to reinvent that process every time. And that's super token exhaustive, it's super slow, and there's no observability or determinism in it. And a lot of our workflows have these small chunks of deterministic steps. And they're also can be a lot more reliable. And so we've made it so that our own agent infrastructure is able to use what we call Vercel Workers and be able to ship those pieces of code. The second thing I'll say also is that a lot of agent processing is best done on a computer. And so as context has exploded and more and more information is available to an agent, giving the agent the accessibility of a local file system to be able to store local thoughts, write short Python scripts, be able to do that processing that's really important when there's so much information.

That in and of itself has been probably one of the most important tools outside of like edit and search, that our internal agents rely on to get work done. And so Vercel Sandboxes have become a critical component in our own agent infrastructure as well.

2:42

SPEAKER_2

No, this is so interesting to say because right now when you're just throwing LLM to any task in a workflow, it's using probabilistic interactions to actually add, which is so token exhaustive. Exactly what you mentioned around, hey, most of your tasks are deterministic. You can just program exactly what it should be done. I feel like you at Notion, you get a really good example of how everyday people are doing their workflows and what they actually want to use agents for. And so can you just break that down? What are some of these like examples of how what you're seeing on the web?

3:13

Sarah Sachs

Yeah, and I will say like users don't say, I want to execute code here. I want to ship a worker. Right, users have outcomes and goals that they want. And part of what we do is like apply taste into when it's appropriate to use those tools. To answer your particular question though, a lot of what our users want are higher and higher level knowledge work tasks. So it used to be that notion was kind of this assistant. And this assistant was where you could say like edit this page, write this thing, make it shorter.

Excuse me. You know, GPT-35 is when we shipped that product. Then we started entering RAG. RAG was the thing, and it was all about answering questions in notion.

It was still an assistant though. I would say in around 2025, we transitioned from notion agents being actual assistants to processes. When agents are processes, that's when they need determinism. That's when they have large amounts of context. That's when they need independent and reliable decision making. And I think the end of that transformation where we are today is having those processes work together in a system. So we're at a higher and higher level of abstraction. When agents are able to work in a system together, there's a lot of state that's shared. There's a lot of universal context that's really important. How you handle that context is critical, and that's where Notion comes to play.

Read the other 13 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000796647858