Simon Last

Simon Last is a Notion cofounder and a builder of its AI systems, from an early Notion AI prototype with Ivan Zhao to the development of Custom Agents. His work includes redesigning how agents interact with Notion and turning evaluation into a workflow agents can help execute.[2][3][1]

In an April 2026 interview with Sarah Sachs, Last described his role as looking ahead at model capabilities and prototyping, while Sachs handled management. Sachs called the fast-changing experimental projects the “Simon Vortex,” with trusted senior engineers moving between prototypes and production work. She credited both Last and Zhao with the openness to discard code and change direction.[1]

Last recalled beginning background-assistant experiments after gaining access to GPT-4 in late 2022. The team designed its own tool-calling framework and tried fine-tuning models to use it over multiple turns. He attributed the failures to insufficient model capability, short context windows, and unreliable behavior, estimating that the eventual agent system had been rebuilt four or five times. Custom Agents required greater reliability because they ran in the background; Sachs separately described repeated work on permissions and the interface administrators needed to understand them.[1]

The early designs exposed a mismatch between Notion’s internal abstractions and the representations models understood. Last recounted starting with a coding agent using JavaScript APIs, then moving to custom XML that mapped losslessly to Notion blocks. The XML was convenient for the team but unfamiliar to the model. They replaced it with Notion-flavored Markdown, accepting that conversion need not preserve every detail, and replaced a complex JSON database-query format with familiar SQL-like queries. His advice was to avoid exposing unnecessary system complexity to the model.[1]

Last also worked on progressive disclosure: letting the agent discover relevant tools instead of placing every tool in its context at once. He said the growing tool set created token overhead and quality problems, with niche tools sometimes disrupting general behavior. In April 2026, he described a latest agent version with more than 100 tools and a shorter prompt; its rollout to Custom Agents was still forthcoming. Sachs explained that tool definitions and evaluation support also let individual teams own their tools rather than compete over one shared prompt.[1]

Last and Sachs recounted delaying the Custom Agents launch by roughly a month to make chat the main interface for setup, testing, and use, with settings alongside it. Last favored giving the agent tools to configure and debug itself, so a user could ask it to explain a failure and revise its instructions in the same conversation. Sachs said three engineers from three different teams worked on the change.[1]

That ability had a boundary: Last explained that background agents began with no permissions and required explicit grants. They could not edit their own permissions. Fixing an agent moved the user into a visible, synchronous administrative interaction, where Sachs said changes required confirmation. Last’s accompanying product principle was to “teach at the top of the class”: sophisticated operators should be able to inspect the tools and understand how the system works.[1]

Last described helping a colleague coordinate more than 30 internal go-to-market agents that researched customers and triaged feedback. They added a manager agent and a shared issue database where the other agents could report problems. In his account, blocked-agent notifications fell from more than 70 a day to about five. He favored composing existing primitives: agents could coordinate through shared databases, and memory could simply be an editable page or database. Direct agent invocation was still awaiting public release at the interview date.[1]

Last described treating the evaluation system as an agent harness—the tools and surrounding workflow through which an agent operates. An agent should be able to download a dataset, run an evaluation, investigate a failure, debug it, and implement a fix, with a human observing the surrounding system. He favored general command-line tools rather than tying the workflow to one coding agent; Sachs emphasized that supervision remained necessary.[1]

His exploratory “software factory” vision extended that approach to agents developing, debugging, reviewing, and maintaining software together. He emphasized readable specifications, strong testing and self-verification, and explicit flows for bugs and pull requests. In his view, engineering increasingly involved observing and maintaining the outer system through which agent work passed—a technical design problem involving approvals, blocked work, and learning mechanisms.[1]

Last favored command-line environments because agents could discover capabilities through help commands, handle long outputs, and debug their tools in the same environment. He also valued MCP for narrower, lightweight agents with tighter tool permissions. He argued that progressive-disclosure problems could be fixed in the harness rather than being inherent to MCP, while command-line access introduced harder questions about credentials and permissions.[1]

After spending substantial time training models, Last argued that frequently changing tools made tool-specific fine-tuning a poor tradeoff. Failures could require a tool fix, better harness design, or verification rather than another training run. He remained interested in company-specific fine-tuning as a possible quality improvement and in having agents replace repeated work with deterministic code: “A lot of the time, the right tool for the job is actually just writing code and not even using agent at all.”[1]

Last’s own working habits reflected his use of long-running agents. He recalled starting model-training experiments before sleep and later doing the same with coding agents, aiming for them still to be working when he woke. He estimated that one continuous agent thread ran for about 17 days and encountered a harness bug.[1]

Last’s collaboration with Zhao predates the AI work. A June 2023 Latent Space account described the pair running low on funding in 2015, laying off the remaining team, discarding their three-year-old app, and rebuilding in a small room in Kyoto, Japan. That work became the beginnings of Notion.[3]

Last and Zhao also built the early Notion AI prototype during the company’s 2022 retreat. Linus Lee later recounted their hack weekend, placing the prototype approximately in late October or early November and explicitly noting that he had joined afterward. Lee identified Last as Notion’s cofounder and CTO in June 2023.[3]

Beyond Notion, Last participated in Imbue’s financing, according to an October 2023 account that listed him among its investors.[6]