Notion's Agentic Evolution: Building the Software Factory
A deep dive into Notion's strategic shift towards custom agents and the 'software factory' concept. The discussion covers the technical hurdles of agent reliability, the importance of model behavior engineering, and the vision for a system of record that caters to both humans and agents.
The Shift Toward Agentic Workflows
Notion is evolving from a collaborative workspace into a sophisticated ecosystem for agents. The core philosophy is to move beyond simple 'wrappers' and instead build a system of record where both humans and agents can coexist and collaborate. This transition is not merely about adding AI features but about redesigning the product to be 'agent-pilled,' where the primary traffic and interface interactions may eventually come from agents rather than humans.
The 'Software Factory' and Technical Rigor
Central to this evolution is the concept of the "Software Factory"—an automated workflow for developing, debugging, and maintaining codebases via coordinating agents. Notion's journey highlights a critical technical insight: the importance of providing models with the environment they actually want (e.g., using SQLite for queries rather than complex JSON). Their approach to reliability focuses on 'Model Behavior Engineering' (MBE), a specialized role that blends linguistics, data science, and prompt engineering to define and measure the 'headroom' of model capabilities.
Strategic Infrastructure and Composability
Notion's strategy involves a pragmatic mix of native integrations and open protocols like the Model Context Protocol (MCP). While MCP is ideal for lightweight, permissioned agents, Notion invests in native builds for high-touch quality areas like search and email. By treating memory as pages and databases—the same primitives humans use—they ensure a seamless loop of memory and execution for agents.
Conclusion
For leadership and investors, the key takeaway is that the value in AI is shifting from the model itself to the orchestration layer and the data primitives. Notion's willingness to rebuild its harness multiple times and its focus on 'demos over memos' demonstrates that in the AI era, agility and the ability to delegate to agents are becoming the primary competitive advantages.
Key insights
-
Coding agents are viewed as the kernel of AGI. The ultimate goal is a 'software factory' where agents can bootstrap their own software, debug, and maintain capabilities autonomously.
Impact: This could drastically reduce the cost of software maintenance and accelerate the pace of feature deployment in enterprise environments.
-
Model Behavior Engineering (MBE) is a distinct career path. It requires a mix of linguistic intuition and technical taste to define 'headroom evals'—tests that models currently fail (e.g., 30% pass rate) to track frontier capabilities.
Impact: Companies will need to shift from traditional software engineering to roles focused on steering and auditing model behavior.
-
The 'Last Exam' philosophy: The most valuable system of record is one where agents can natively interact with the same primitives (pages, databases) that humans use, creating a shared memory space.
Impact: Creates high switching costs and extreme lock-in by making the system of record indispensable for both human and agent productivity.
-
Progressive disclosure of tools is essential for scaling agents. Providing a model with 100+ tools simultaneously degrades quality; agents must search for and 'discover' the right tool for the task.
Impact: Enables the deployment of massive tool libraries without compromising the reasoning capabilities of the underlying LLM.
-
Value alignment in pricing: Using language models for deterministic tasks is wasteful. Pricing and architecture should shift toward executing code/CLIs for deterministic actions to optimize cost and latency.
Impact: Increases profitability and accessibility by reducing token waste for simple, repeatable tasks.
Action items
-
Transition from 'few-shot' prompting to goal-driven tool definitions and robust evaluation frameworks. This allows tool ownership to be distributed across different product teams rather than a centralized 'prompt' file.
Impact: Increases engineering velocity and prevents the 'center of excellence' bottleneck in AI development.
-
Implement a 'demos over memos' culture and a high-tolerance approach to deleting code. In a rapidly shifting AI landscape, the ability to rebuild the core harness multiple times is a competitive advantage.
Impact: Prevents technical debt from anchoring the product to outdated model capabilities.
-
Build 'agentic find' capabilities by optimizing retrieval for agents rather than humans. This includes fanning out parallel queries and focusing on top-k retrieval over traditional click-through ranking.
Impact: Significantly improves the accuracy of agent-led information retrieval in large-scale enterprise knowledge bases.
Quotes
“I think one thing that's becoming more clear is I think the coding agents are the kernel of AGI. Everything is a coding agent.”
“Our objective is to make it so that the whole product org is building for agents.”
“If you're just pressing against model capabilities versus not exposing the model to the right information... that in and of itself is a skill of intuition.”