Notion treats its AI harness as inherently perishable, rewriting it roughly every six months to match current model capability rather than accumulating technical debt around an aging design.
Last calls this a 'running joke' internally but says it is a deliberate discipline: companies that build a harness once and stick with it fall behind, because the right system design depends on what the current generation of models is actually good at. The rewrite interval itself has been shrinking as model progress accelerates, and coding agents now make each rewrite faster to execute.
engineering-org-transformation
Building agent-facing APIs is a distinct design problem from building human-facing APIs, and Notion had to redesign both its page format and database access layer once agents became a real customer.
Notion's original block-based JSON API was, in Last's words, 'crazy verbose' and unusable for agents. The team built a markdown dialect that reads like normal markdown but encodes all Notion block types, and moved database access to a SQLite-like query interface. The design process was empirical: try things, notice what burns too many tokens or underperforms, and reason from first principles about what patterns are already well-represented in model training data.
agent-native-product-design
Retrieval quality across heterogeneous data sources like Slack and Google Drive is less about the AI model and more about domain-specific craft: chunking strategy, per-source query tuning, and constant empirical iteration.
Last says Notion was 'kind of baffled' that most companies had not solved their own search well, and attributes Notion's edge to being 'AI-pilled' plus a lot of unglamorous attention to detail - actually trying queries against each data source, since Slack and Google Drive need fundamentally different retrieval treatment. He notes that with embeddings, the user's own organizational structure inside Notion matters far less than it used to, since the AI just needs a retrievable text snippet with the right context.
retrieval-and-indexing-craft
Coding agents have shifted Last's own role from writer of code to designer-and-verifier of end-to-end tasks, and he says he has not personally written code since mid-2025.
He describes the progression as tab-autocomplete, then agent-assisted small edits, then full end-to-end delegation where he defines the task and its verification and monitors for the agent going 'off the rails.' He frames this explicitly as becoming 'the agent manager instead of the coder,' and says this shift lets him be more ambitious about what he attempts to build.
engineering-org-transformation
Aggressive agent adoption inside Notion has made engineering feel messier but has also raised the ceiling on individual output, with the gap between the most and least tool-fluent engineers widening sharply.
Last says the minimum competence bar for engineers has not moved much, but the maximum has 'extremely increased' - a single engineer willing to fully harness agents can be '100 or 1,000x' more productive. He cites Notion's design team building an entire internal 'design playground' repo with its own agent, letting designers ship deployed high-fidelity prototypes instead of static mocks.
engineering-org-transformation
PRs written by coding agents tend to be larger and more complex, which raises review burden, but they also come with much better test coverage because agents can be held to a higher testing bar than a human would set for themselves.
Last says he never ships a PR anymore that has not been fully agent-tested, and that the real skill is not 'vibe coding' by describing what you want, but carefully specifying what change you're making and how it can be verified and deployed safely, then enlisting the agent to execute that plan.
engineering-org-transformation
Notion's personal-agent and custom-agent products only became viable on roughly the third or fourth attempt, after multiple years of a general assistant vision that 'always sort of works, but then wasn't that useful.'
Notion shipped a narrow AI writer first (fastest to build, no retrieval needed), then Q&A grounded in a semantic index (GA'd October 2023), and only launched a fully agentic personal assistant with workspace-wide tool access in August/September of the prior year, followed by custom background agents the week of this recording.
agent-native-product-design
Notion positions itself as a model-agnostic 'Switzerland' layer, deliberately avoiding lock-in to any single AI lab and actively integrating multiple Chinese open-source models because they are notably cheaper and increasingly competitive with frontier models.
Last says customers 'don't want to be locked into a certain lab's model' given how often the leaderboard reshuffles month to month, and that Notion's role is to make it easy to access and switch between all the best available models, including open-source options, rather than betting on one provider.
model-agnostic-platform-strategy
Notion's custom agents are designed to bootstrap their own capabilities from a minimal starting kernel, including writing and deploying new integrations the platform does not natively support.
Last frames coding agents as 'the kernel of AGI' because code is a general-purpose primitive for deterministic logic; applied to a knowledge-work agent, this means an agent that hits a missing integration can, in principle, build that integration itself rather than waiting on Notion's engineering team.
agent-native-product-design
Last's own daily workflow depends on running many autonomous coding agents continuously, deliberately queuing enough work at night that agents are still running when he wakes up; his personal record is a single agent run of 13 days straight.
He describes an explicit goal of maximizing concurrent agent runtime - 'my whole goal these days is essentially to just have as many running as possible, and to run them all the time' - and admits to checking on long-running agents in the middle of the night out of nervousness that they might have stalled.
autonomous-personal-agents
Last's personal email-triage custom agent shows a repeatable pattern for training an autonomous agent: let it propose actions, correct it for a couple of weeks, then remove the approval step entirely once trust is established.
He gave the agent access to his email and a blank memory page, had it propose which emails to archive, corrected its judgment for the first couple of days, and after a couple of weeks dropped approval entirely - it now archives everything he does not need to see. He describes this as having 'completely solved' his email problem since he barely used email for work to begin with (most work communication happens in Slack).
autonomous-personal-agents
Even as Notion becomes agent-first, the same core primitives - documents and databases - remain necessary; what changed is that a new primitive (the agent itself, and how it interacts with pages and databases) had to be added on top.
Last says agents still love writing markdown documents and still need structured databases for coordination - if you're running a swarm of 100 background coding agents, you don't want 100 separate chat threads, you want a Kanban board. The fundamental goal shifted from 'best tool for humans to do work directly' to 'best tool for humans to manage agents doing the work,' but the underlying data model held up.
agent-native-product-design
Companies
Notion - Last's company; the episode traces its shift from a writing/knowledge-base tool to a platform where humans and AI agents collaborate, including a personal agent for every user and custom background agents launched the week of recording.
OpenAI - Last describes getting early access to GPT-4 (via a proto-ChatGPT interface) at a company offsite in Mexico in 2022, which convinced him and co-founder Ivan Zhao that AI-native features were immediately viable.
Slack - One of the external sources Notion indexes into its semantic search so its Q&A and agents can answer questions grounded across a user's whole workspace, not just Notion pages.
Google Drive - Second external source Notion indexes alongside Slack for cross-workspace semantic search.
Anthropic - Implied via Last's daily use of Claude Code as his primary coding-agent CLI tool, alongside Codex.
Techniques and frameworks
Six-month harness rewrite cadence - Notion's AI team rewrites its agent harness roughly every six months because model capability moves fast enough that a harness optimized for last generation's models becomes suboptimal; the rewrite interval has been shrinking as progress accelerates.
Agent-native API design - Notion redesigned its block/page API into a markdown-like dialect and moved databases to a SQLite-style query interface specifically because the original human/developer-oriented JSON API was too verbose and awkward for agents to use well.
Approval-mode to autonomous handoff - Last's pattern for every custom agent he builds: run it in supervised/approval mode while it learns his preferences from corrections, then drop the approval step once he trusts it, checking back in only when it breaks.
Outer-loop verification - Last's current engineering workflow: design an end-to-end task and its verification criteria, hand execution to a coding agent, and act only as the outer verifier checking correctness and catching derailment, rather than writing code directly.
Summary
Sarah Guo interviews Notion co-founder Simon Last about how the company's AI strategy evolved from an early writing assistant into a full agent platform, and how that shift has reorganized Notion's own engineering culture. Last traces the origin story to a 2022 company offsite in Mexico where he and co-founder Ivan Zhao got early access to GPT-4 through a proto-ChatGPT interface and immediately saw two things: the model could follow complicated instructions, and its breadth of world knowledge was startling. That conviction produced parallel short-term and long-term bets - a quick-to-ship AI writer, and a much harder, multi-year push toward a general assistant with access to all of Notion's own tools (databases, documents, search).
A large part of the conversation is about the unglamorous engineering work behind retrieval. Notion indexes not just its own pages but external sources like Slack and Google Drive, and Last is candid that most companies, including large well-resourced ones, have not solved search over their own data well - something he says "kind of baffled" the team. He attributes Notion's edge to being an early, aggressive adopter of the technology combined with deep craft: constant empirical query testing, source-specific chunking and retrieval tuning, and a willingness to treat retrieval as a living system rather than a one-time integration.
The episode's most concrete organizational claim is that Notion rewrites its AI agent harness roughly every six months, and that this interval keeps shrinking as model capability accelerates. Last frames this not as wasted work but as a required discipline: teams that build a harness once and stop iterating fall behind, because the right system design is a function of what current models are actually good at. This same acceleration has changed his own daily practice - he says he has not personally written code since mid-2025, describing a shift from typing code, to tab-completing it, to delegating end-to-end tasks to coding agents while he acts purely as the "outer verifier," checking correctness and catching derailment.
Last also details how Notion redesigned its APIs specifically for agents as a new class of customer: the original JSON block format was too verbose for models to use well, so the team built an agent-friendly markdown dialect for pages and a SQLite-style query layer for databases. He connects this to a broader platform bet - Notion aims to be a model-agnostic "Switzerland," giving customers easy access to and easy switching between all major labs' models (including several competitive, cheaper Chinese open-source models) rather than locking into one provider.
The back half of the conversation turns personal: Last describes running many autonomous coding agents continuously, deliberately loading enough overnight work that agents are still running when he wakes, with a personal record of one agent running 13 days straight. He walks through his own custom agents in daily use, most notably an email-triage agent he trained by correcting its proposed archive decisions for a couple of weeks before removing approval entirely, and a Slack bug-triage agent that learned hundreds of routing rules over time. He closes by describing Notion's underlying philosophy: even as the product becomes agent-first, the core primitives of documents and databases remain essential - what changed is that a new primitive, the agent and its relationship to pages and databases, had to be added, and the company's goal shifted from being the best tool for humans doing work directly to being the best tool for humans managing agents that do the work.
Notable Quotes
"I think of coding agents as like the kernel of AGI. AGI will be a coding agent." - Simon Last
"You're not just vibe coding by saying the thing you want... you're sort of thinking carefully about what is the change I'm trying to make, and how can it be verified and how can it be deployed safely." - Simon Last
"My whole goal these days is essentially to just have as many [agents] running as possible, and to run them all the time." - Simon Last
"I'm now like the agent manager instead of the coder." - Simon Last
"Before AI, our goal is to create the best tool for humans to directly perform their work. And then now the goal is to create the best tool for humans to manage agents to do the work for them." - Simon Last