Hey everyone,
I’m currently building a personal AI agent for a client that should support them with calendar, email, and project management (in Notion). I’d like to share my current setup and get your thoughts on how you would approach a project like this from scratch, and what pitfalls you see.
Here is my current workflow:
Right now I have a main orchestrator (Main Agent) that receives incoming messages (text). This orchestrator then calls different sub-agents depending on the intent. Each sub-agent is responsible for a specific tool or function and then executes the corresponding action.
The sub-agents I currently have are:
-
Email Agent → creates Gmail drafts
-
Calendar Agent → creates/updates/deletes Google Calendar events
-
Notion sub-agents →
-
Research Agent (fetches customer and project info from Notion and give me details about it)
-
Meeting Agent (can create and retrieve meetings including content)
-
Task Agent (can create and retrieve tasks)
-
Notes/Thoughts Agent (can create and retrieve notes / thoughts)
-
So the workflow is: Chat → Main Orchestrator → Sub-Agent → Tool (Gmail, GCal, Notion) → result goes back and is logged into Google Sheet.
The client’s main goal is to have an AI agent that acts as a true all-round personal assistant in their daily business. It should support them at any time, whether they are at the desk, on the move, or even while driving.
They want the agent to function like a “second brain” that remembers everything: customer updates, project tasks, meeting notes, emails, offers, ideas… The agent should always have full context, understand the ongoing state of each project, and be able to provide meaningful updates in natural language, not just raw database entries.
Notion should be the central hub where all information is stored and structured, but the agent should be able to surface that information conversationally, so the client can always ask, “What’s going on with project X?” and receive a clear, contextual answer. In short, the client wants a digital memory and assistant combined - a system that knows the entire history, manages project workflows, and supports them across all aspects of their business in real time.
My questions: is it a good approach to work with one main orchestrator and multiple sub-agents, or would you set it up differently to reduce error risks?
The main problem I see right now is that the setup feels too error-prone. The orchestrator sometimes triggers the wrong sub-agent, or passes the wrong output along, and the sub-agents themselves don’t always behave consistently. Everything has to be defined very precisely for the flow to work reliably, and at the moment it’s still not running in the smooth, intuitive way the client expects.
I’d love to hear how you in the community would approach this. Would you go with an orchestrator + sub-agents like I have now? Or would you prefer another pattern? And what best practices do you use to avoid common pitfalls?
Thanks a lot for your input!