A new shape for work
Notion argues that teams get little from AI until they build a shared context system that people and agents work in together, then shared "factories" that run recurring tasks for everyone. It's a sales-flavored field guide, so read the product pitch with that in mind.
- Chat-based AI is stateless and single-player; value compounds only when agents read from and write to a central, well-structured knowledge store.
- Shared "factories" (specific, shared, scrutinized agent systems) beat everyone hand-rolling personal agents, because fixed setup costs get paid once and improvements compound.
- The article's concrete build process: notice a repeatable task, plan with a simple version, run it in shadow mode first, then improve it against written evals.
Notion's field guide, "A New Shape for Work," argues that most teams using AI aren't getting much out of it, and the reason is a missing "context system": a shared, organized body of information that both people and their agents can read from and write to. The authors compare unstructured AI use to adding a power boost to a bicycle with one gear. The pedals spin faster, but the bike goes nowhere new. Their symptoms list is familiar to anyone who has tried this: information scattered across tools, agents that can't find things or lack background a teammate would have, one-off chatbot demos with no durable process behind them, and hundreds of individually built agents nobody can audit for access or risk. Each of these, they say, traces back to the same gap.
The proposed fix has two parts. First, build a shared context system (the guide discusses structured object storage, an ontology, a fractal structure, syncs, meetings, permissions, validation, and ways to counter "context rot," where stored information goes stale). Second, build shared AI systems they call "factories," owned by a small group and used by everyone else. A factory can be as simple as a shared to-do list tied to an agent, or as elaborate as automation triggered when a task is created or an email arrives. The guide describes a four-step loop for building one: notice a recurring problem, plan, run it, then improve it. The pitch is specialization and trade applied to AI work, so people spend their time on what only they can do.
Practically, the guide is a vendor's playbook for its own product. Notion is explicit that its goal is to make Notion the natural home for this approach, though it says the strategy works with any model provider or toolset. The ideas are plausible and the diagnosis of disorganized information is hard to argue with. What's less clear from the excerpt is how much of the benefit depends on a team already having a clean, well-maintained knowledge base, which is often the hard part. Teams weighing this should expect the real cost to be the upfront work of organizing their information, not the agent setup itself, and the guide's suggestion to start with one workflow is the sensible way in.