- Notion is evolving into an agent orchestration platform, allowing seamless collaboration between humans and AI agents within their workspace.
- They have integrated "managed agents" like Claude to handle complex, long-running tasks, making advanced AI capabilities plug-and-play for users.
- This system enables non-technical users to automate multi-step workflows, such as client onboarding, by leveraging Notion's rich context and powerful AI agents.
How Notion built with Claude Managed Agents
- Notion aims to be an agent orchestration platform, bringing together the right AI agents for specific jobs and managing associated workflows.
- Integrating "managed agents" like Claude simplifies the use of powerful AI for complex tasks, avoiding the need for extensive engineering efforts.
- Managed agents are designed to handle long-running tasks (e.g., 20 minutes to an hour), effectively managing memory and ensuring high-quality outputs over time.
- Notion's custom agents can access full context from internal databases and task boards, allowing AI to make informed decisions and generate relevant outputs.
- Users can kick off multiple "agent threads" concurrently within Notion, which feed into a Claude session for parallel processing.
- The platform provides a mirrored view of running Claude sessions, enabling direct interaction with the AI and offering insights into development traces for improvement.
- The managed agent product allows customers to automate workflows and iterate on AI-generated content directly within Notion.
agent orchestration platform — A system designed to manage, coordinate, and integrate multiple AI agents and their interactions with human workflows.
managed agents — Pre-configured and maintained AI agents (like Claude) integrated into a platform, offering plug-and-play functionality without requiring users to handle underlying infrastructure.
Claude — A specific large language model (LLM) developed by Anthropic, known for its strong performance on complex and long-running tasks.
long-running tasks — Computational processes that require extended periods to complete, potentially minutes or hours, and often involve managing memory and maintaining quality over time.
API — Application Programming Interface; a set of rules and protocols that allows different software applications to communicate and exchange data.
agent threads — Individual parallel execution paths or instances of an AI agent performing specific sub-tasks within a larger workflow.
Claude session — A dedicated, interactive instance of the Claude large language model running to process a set of tasks or a conversation.
traces — Detailed logs or records of the steps, inputs, and outputs of an AI agent's execution, used for debugging, monitoring, and performance improvement.
context — All the relevant information, data, and background available to an AI agent, allowing it to understand the current situation and make informed decisions.
Hi, I'm Eric. I'm a product manager at Notion, and I work on our agents and agent orchestration. I had 30 tasks to make a prototype. I took all of that and just drag it to start, and it was like a Claude running wild. I was like, oh, I don't need to watch this. So I got like a snack. I came back and like, all the prototypes were made. That was sort of the unlock. We want Notion to be the agent orchestration platform. Notion is such a good place for human and agent collaboration. You bring the right agent together for the job, and we help to manage all of your workflows. We were really excited about bringing managed agents into Notion, because Claude is really good at complex, long running tasks. If you work to roll it up yourself, it's like a mega brain engineering effort. You need like a PhD in all of this to make it work. And so the managed agent product was great because it makes it really plug and play to bring in Claude. It runs the sessions in the Claude. We just pull in the API and it works for our customers. What that means is that they can kick off a ton of jobs in Notion. I'd love to show a use case of how we use managed agents to build agent orchestration in Notion. So this example is something that a lot of our non-technical users do, which is they'll onboard clients. Let's take a fictitious client called Harbor and Pine. They are a lifestyle brand and now here are all the action items. We have within Notions a custom agent, and we've basically given it all of the client databases and task boards. So it has the full context to onboard, and it helps us in the onboarding workflows for clients behind the scenes. This agent has Claude as well. It is pulling in all the contexts within notion to go and generate this. And so what I want to do is call upon the client onboarding manager and then take these action items, turn them into tasks so you can see them. They're all in this task board. And this is what I like about it. Like I can take all of this and just dump it here. And you can see on the side that it kicks off a bunch of agent threads. And that all basically feeds into Claude. And then Claude will then kick off a Claude session on that. And so now you can, talk to that Claude session within Notion here, which mirrors the actual session that was kicked off. So if we go into the Claude platform so you can see that this session is running for Claude manage agents, it kind of gives you a different view of what's happening. And it's useful for us on the development side. To go and see like what's happening, but also actually to feed all of these traces to go and improve our agent. Having a harness that can do long running tasks is really essential. You might need to run it for 20 minutes an hour. That ability to continue to run it, to manage memory, to have high quality outputs over time, is a layer that's super critical on top of the model itself. A lot of the tasks are completed already. This is what a example homepage could look like, and it is pulling in all the contexts within Notion to go and generate this and if we wanted to iterate on that, we can talk to Claude directly here. That's how we used managed agents in Notion to help ourselves and help our customers. So the managed agent product is like a playground for me. And I just really cool to be able to kick off all these jobs at the same time. I just love building cool s#!t.
TL;DR
- Notion is evolving into an agent orchestration platform, allowing seamless collaboration between humans and AI agents within their workspace.
- They have integrated "managed agents" like Claude to handle complex, long-running tasks, making advanced AI capabilities plug-and-play for users.
- This system enables non-technical users to automate multi-step workflows, such as client onboarding, by leveraging Notion's rich context and powerful AI agents.
Takeaways
- Notion aims to be an agent orchestration platform, bringing together the right AI agents for specific jobs and managing associated workflows.
- Integrating "managed agents" like Claude simplifies the use of powerful AI for complex tasks, avoiding the need for extensive engineering efforts.
- Managed agents are designed to handle long-running tasks (e.g., 20 minutes to an hour), effectively managing memory and ensuring high-quality outputs over time.
- Notion's custom agents can access full context from internal databases and task boards, allowing AI to make informed decisions and generate relevant outputs.
- Users can kick off multiple "agent threads" concurrently within Notion, which feed into a Claude session for parallel processing.
- The platform provides a mirrored view of running Claude sessions, enabling direct interaction with the AI and offering insights into development traces for improvement.
- The managed agent product allows customers to automate workflows and iterate on AI-generated content directly within Notion.
Vocabulary
agent orchestration platform — A system designed to manage, coordinate, and integrate multiple AI agents and their interactions with human workflows.
managed agents — Pre-configured and maintained AI agents (like Claude) integrated into a platform, offering plug-and-play functionality without requiring users to handle underlying infrastructure.
Claude — A specific large language model (LLM) developed by Anthropic, known for its strong performance on complex and long-running tasks.
long-running tasks — Computational processes that require extended periods to complete, potentially minutes or hours, and often involve managing memory and maintaining quality over time.
API — Application Programming Interface; a set of rules and protocols that allows different software applications to communicate and exchange data.
agent threads — Individual parallel execution paths or instances of an AI agent performing specific sub-tasks within a larger workflow.
Claude session — A dedicated, interactive instance of the Claude large language model running to process a set of tasks or a conversation.
traces — Detailed logs or records of the steps, inputs, and outputs of an AI agent's execution, used for debugging, monitoring, and performance improvement.
context — All the relevant information, data, and background available to an AI agent, allowing it to understand the current situation and make informed decisions.
Transcript
Hi, I'm Eric. I'm a product manager at Notion, and I work on our agents and agent orchestration. I had 30 tasks to make a prototype. I took all of that and just drag it to start, and it was like a Claude running wild. I was like, oh, I don't need to watch this. So I got like a snack. I came back and like, all the prototypes were made. That was sort of the unlock. We want Notion to be the agent orchestration platform. Notion is such a good place for human and agent collaboration. You bring the right agent together for the job, and we help to manage all of your workflows. We were really excited about bringing managed agents into Notion, because Claude is really good at complex, long running tasks. If you work to roll it up yourself, it's like a mega brain engineering effort. You need like a PhD in all of this to make it work. And so the managed agent product was great because it makes it really plug and play to bring in Claude. It runs the sessions in the Claude. We just pull in the API and it works for our customers. What that means is that they can kick off a ton of jobs in Notion. I'd love to show a use case of how we use managed agents to build agent orchestration in Notion. So this example is something that a lot of our non-technical users do, which is they'll onboard clients. Let's take a fictitious client called Harbor and Pine. They are a lifestyle brand and now here are all the action items. We have within Notions a custom agent, and we've basically given it all of the client databases and task boards. So it has the full context to onboard, and it helps us in the onboarding workflows for clients behind the scenes. This agent has Claude as well. It is pulling in all the contexts within notion to go and generate this. And so what I want to do is call upon the client onboarding manager and then take these action items, turn them into tasks so you can see them. They're all in this task board. And this is what I like about it. Like I can take all of this and just dump it here. And you can see on the side that it kicks off a bunch of agent threads. And that all basically feeds into Claude. And then Claude will then kick off a Claude session on that. And so now you can, talk to that Claude session within Notion here, which mirrors the actual session that was kicked off. So if we go into the Claude platform so you can see that this session is running for Claude manage agents, it kind of gives you a different view of what's happening. And it's useful for us on the development side. To go and see like what's happening, but also actually to feed all of these traces to go and improve our agent. Having a harness that can do long running tasks is really essential. You might need to run it for 20 minutes an hour. That ability to continue to run it, to manage memory, to have high quality outputs over time, is a layer that's super critical on top of the model itself. A lot of the tasks are completed already. This is what a example homepage could look like, and it is pulling in all the contexts within Notion to go and generate this and if we wanted to iterate on that, we can talk to Claude directly here. That's how we used managed agents in Notion to help ourselves and help our customers. So the managed agent product is like a playground for me. And I just really cool to be able to kick off all these jobs at the same time. I just love building cool s#!t.