Notion review Custom Agents public beta: AI office shifts from "question and answer" to "entrustment"

On June 15, Notion co-founder Akshay Kothari posted a review of the gains made during the public beta of Custom Agents. The value of this review does not lie in what new features it talks about, but in that it clarifies a shift in product philosophy: Notion's Custom Agents are not "smarter chatbots", but "AI colleagues" - taking repetitive tasks away from users and allowing users to focus on the work that "needs you (people)".

Core Concept: Delegation rather than dialogue

The core insight of the review is "task delegation (delegation)" rather than "dialogue". The difference between the two is: dialogue means you ask a question and it answers a question; delegation means you set roles, goals and permissions for the Agent, and the Agent executes independently in the Notion workspace - organizes data, generates documents, maintains the database - and delivers the results after completion. This requires the Agent to have reliable task understanding, workspace operation and result verification capabilities. In other words, the user changes from "operator" to "manager".

Complete stack from application to model layer

Custom Agents is not an isolated function, it is being woven into a larger web by Notion: Custom Agents (execution layer) + Developer Platform (developer platform, released on June 1) + ZeroEntropy (knowledge work-specific model team, merged on July 1). Together, these three pieces of the puzzle outline Notion's ambition to build a complete stack of "application layer + agent layer + model layer" - to control both execution and the underlying model.

From an industry perspective, Custom Agents represent a key paradigm shift in AI office products from "question-and-answer assistants" to "entrusted AI colleagues." The significance of this change is that when AI can autonomously complete an entire workflow and deliver results, the imagination of "AI office" will be upgraded from "efficiency improvement tool" to "manpower replacement" - and this is the core logic that companies are willing to pay for. Notion is using public beta experience to prove that this road is feasible. For domestic collaborative office AI (Feishu, DingTalk, etc.), the reference value of this review lies in the execution details: role setting, authority boundaries, result verification - the success or failure of "entrusted AI colleagues" is precisely hidden in these seemingly trivial engineering details.

Several directions worth tracking in the future:

1. Success rate of delegated tasks: completion quality and exception handling of autonomous execution.
2. Design of permission model: Agent’s access boundaries and approval flow in the workspace.
3. Path from public beta to official launch: Product improvement based on review experience.
4. The "delegation" of domestic office AI: Who will let AI move from "question and answer" to "autonomous execution" first?