CHAPTERS
From “30 tasks” to instant prototypes: the personal unlock
Eric, a Notion product manager, describes a breakthrough moment: dumping a long list of tasks into an agent workflow and returning to find prototypes already completed. That experience reframes what’s possible when agents can execute autonomously at scale.
- •Eric’s role: agents and agent orchestration at Notion
- •A rapid prototype sprint with ~30 tasks as the catalyst
- •Hands-off execution: “Claude running wild” while he steps away
- •Outcome: multiple prototypes produced without constant supervision
Notion’s vision: an agent orchestration platform for human–agent collaboration
Eric explains the broader product direction: Notion as the place where people and specialized agents work together. The goal is to assemble the right agents for the job and manage workflows end-to-end inside Notion.
- •Positioning Notion as an agent orchestration platform
- •Emphasis on human + agent collaboration in one workspace
- •“Bring the right agent together for the job” framing
- •Workflow management as the core value layer
Why Claude Managed Agents: long-running, complex tasks without the “mega brain” build
He highlights why Managed Agents matter: building robust long-horizon agent systems in-house is costly and deeply technical. Claude Managed Agents reduce the engineering burden by making advanced agent capabilities more plug-and-play.
- •Claude’s strength: complex, long-running tasks
- •DIY agent infra described as a major engineering effort
- •Managed Agents as a simpler, integrated product surface
- •Lower barrier for teams that don’t have deep agent-systems expertise
How Notion integrates Managed Agents: cloud sessions + API simplicity
Eric outlines the integration model: Claude sessions run in the cloud while Notion connects via API. For customers, this enables launching many jobs directly from Notion without managing underlying infrastructure.
- •Claude runs sessions in the cloud
- •Notion “pulls in the API” to connect
- •Customers can kick off many concurrent jobs in Notion
- •Infrastructure abstraction: less operational overhead for users
Use case setup: client onboarding for non-technical users
He introduces a common workflow—onboarding clients—chosen specifically because non-technical teams do it frequently. A fictional client, Harbor and Pine, serves as the demo scenario.
- •Focus on a real, repeatable business process: onboarding
- •Designed to work for non-technical users
- •Demo client: Harbor and Pine (lifestyle brand)
- •Sets the stage for task generation and workflow automation
The custom Notion onboarding agent: grounded in databases and task boards
Notion uses a custom agent that has access to relevant client databases and task boards, giving it the context needed to execute onboarding work. Claude powers the agent behind the scenes to synthesize that context into actionable outputs.
- •A custom agent with access to client databases/task boards
- •Full-context grounding inside Notion
- •Claude as the underlying reasoning/execution layer
- •Generates onboarding artifacts and action items from workspace context
Turning action items into structured tasks: dumping inputs and letting threads run
Eric demonstrates taking a set of action items and converting them into tasks on a Notion task board. The system fans out into multiple agent threads that feed into Claude to execute work in parallel.
- •Action items are transformed into tasks automatically
- •Inputs can be “dumped” in without careful formatting
- •Multiple agent threads kick off concurrently
- •Parallel execution flows into Claude for processing
In-Notion conversation view: chatting with the same Claude session
The demo shows that users can talk to the running Claude session directly inside Notion. This interface mirrors the session created by Managed Agents, enabling iteration and follow-up without leaving the workspace.
- •Chat interface embedded in Notion
- •Mirrors the underlying Claude session that was kicked off
- •Supports iterative refinement and follow-up requests
- •Keeps execution and review in the same place
Developer visibility in the Claude platform: session view, traces, and debugging
Eric switches to the Claude platform to show the session’s execution view. This is valuable for developers to understand what’s happening and to collect traces that help improve the agent’s behavior over time.
- •Claude platform provides an alternate execution view
- •Useful for development: inspecting what’s happening
- •Traces can be used to improve and tune the agent
- •Supports operational debugging and quality iteration
Why “managed” matters: long-running harness, memory, and sustained quality
Eric emphasizes that durable agent execution requires more than a model—it needs a runtime layer that can persist for minutes or hours, manage memory, and maintain output quality across time. This “harness” is positioned as critical infrastructure for real workflows.
- •Long-running jobs may take 20 minutes to an hour
- •Need for persistence/continuation beyond a single prompt
- •Memory management as a key capability
- •Quality over time depends on orchestration/runtime, not just the model
Results and iteration: generating an onboarding homepage and refining via chat
With many tasks completed, Eric shows an example of an onboarding homepage generated from Notion context. He notes that iteration is straightforward: users can continue the conversation with Claude to refine outputs.
- •Tasks complete quickly and visibly in the workflow
- •Example output: a generated homepage for the client
- •Outputs are grounded in Notion context
- •Iteration loop: refine directly by chatting with Claude
Wrap-up: Managed Agents as a scalable playground for running many jobs at once
Eric concludes by describing Managed Agents as a “playground” that makes it easy to launch many simultaneous jobs. The closing reinforces his enthusiasm for building with this multi-job orchestration capability.
- •Managed Agents enable many concurrent workflows
- •Feels like a flexible sandbox for experimentation
- •Scales task execution without constant oversight
- •Personal enthusiasm for building with agent orchestration
