The Agent's Workspace

Notion began as the best tool for humans to perform their work directly. Now, in Simon Last’s words, the goal is to create the best tool for humans to manage agents doing the work for them. That is not an incremental upgrade. It is a different product for a different customer—one that happens to share an address with the old customer.

Ref. E50A-A

What makes this shift remarkable is not that Notion bolted an AI chatbot onto a productivity app. It is that Last and his co-founder Ivan discovered something more fundamental: the same primitives that made Notion useful for humans—documents, databases, kanban boards—turn out to be exactly what agents need . Agents love to write markdown documents. If you are managing a hundred background coding agents, you do not want a hundred chat threads—you want a kanban board. Structured data, unstructured narrative, and coordination surfaces are not human affordances layered on top of a machine substrate. They are the substrate itself.

Ref. FBD8-B

This is the deepest claim in the interview, though Last never quite frames it this way. If he is right, then the winning knowledge-work platform in the agent era is not the one that builds the best AI—it is the one whose existing information architecture already speaks the language agents think in. And Notion, almost by accident of its original design philosophy, may have built exactly that.

Ref. 76DF-C

Last introduces a framing that deserves to be carried far beyond Notion: the agent is a new customer. In the past, Notion designed its product to be convenient for humans, then designed APIs to be convenient for humans writing code against those APIs. Now there is a third customer—the agent—whose needs are related to but distinct from both. The agent does not browse a UI. It does not think in REST endpoints. It thinks in tokens, and it has priors shaped by its training data.

Ref. B64F-D

A recurring thread in the conversation is what transfers from the pre-AI era and what does not. Last is emphatic: almost all of Notion’s core primitives still matter. Documents remain the natural container for unstructured knowledge. Databases remain the natural container for structured data. The coordination surfaces—kanban boards, calendars, linked databases—remain essential because coordination is coordination regardless of whether the workers are human or artificial.

Ref. D480-E

This is a structural insight about platform design. The platforms that win in the agent era will not be the ones that build agents as add-on features. They will be the ones whose architecture treats agents as peers—entities with identities, permissions, and collaborative presence—within an information environment designed for shared work.

Ref. F993-F

What is new is the concept of the agent itself as a first-class entity in the workspace. An agent needs representation: a name, permissions, access boundaries, a memory surface. Last describes creating custom agents that have access to their own databases, can file tasks, can respond in Slack channels, can operate on schedules. The agent is not a feature of Notion—it is a user of Notion, with its own identity and its own capabilities, operating within the same collaborative environment as human users.

Ref. F8B5-G

Last reveals what may be the most counterintuitive engineering practice in the interview: Notion rewrites its AI harness—the system that orchestrates model interactions, retrieval, tool use, and agent behavior—approximately every six months. And the rewrite cycle is accelerating.

Ref. F453-H

“You really do have to be keenly aware of what the current state of the models and the technology is and then design the harness deeply around that. It basically means you have to rewrite it every six months.”

Ref. 5A04-I

There is a deeper principle at work. Last says he finds the rewrite process fun —you get to restart and rethink. This attitude is not incidental to Notion’s success with AI; it is constitutive of it. The willingness to throw away working code in favor of a better architecture requires a specific organizational culture: one that values learning velocity over sunk-cost preservation, and that treats each generation of AI infrastructure as a hypothesis to be tested rather than an investment to be protected.

Ref. F37C-J

The six-month rewrite cycle works because AI capabilities are moving faster than traditional software iteration allows. Each new model generation changes the optimal architecture so substantially that incremental patching accumulates more debt than starting fresh. This works because the rewrite gets faster each time (you understand the problem space better and have better tools). It breaks if your organization treats code as an asset to preserve rather than a hypothesis to test.

Ref. E67F-K

The word craft appears repeatedly. Getting retrieval right, in Last’s telling, requires an unusual combination of empirical rigor and artisanal attention. Each data source is different. You cannot apply a one-size-fits-all chunking and retrieval strategy across Slack messages (short, conversational, context-dependent) and Google Drive documents (long, structured, self-contained). The team had to develop source-specific tuning—trying queries, using the system daily, constantly iterating on how retrieval works for each integration.

Ref. CEE4-L

Retrieval quality is not a model problem—it is an engineering-craft problem. It requires trying a bunch of different queries, actually using it every day, and constantly iterating. There is a little bit of AI-pilled savviness that matters, but most of it is attention to detail and love for the work.

Ref. F688-M

Notion’s original API, for example, uses a verbose JSON format for blocks that was adequate for human developers but catastrophic for agent consumption—too many tokens, too much structural noise. The team redesigned their agent-facing interfaces around two insights. First, they created a markdown dialect

Ref. F4AA-N

The Dual API Challenge

Ref. CDDA-P

Last describes a clear progression through multiple eras of AI-assisted development at Notion. The tab autocomplete era was useful but incremental. The insertion and rewriting era added convenience. But the real shift came when coding agents started working—Last pinpoints his adoption of Claude Code around April 2024 as the inflection.

Ref. 6D6B-R