The Synthesis Layer - Where Raw Outputs Become Coherent
The Synthesis Layer - Where Raw Outputs Become Coherent
An AI agent talks to your email, your calendar, your CRM, your file system, and your browser. Each tool returns raw data. The email API gives you message objects. The calendar gives you event blocks. The CRM gives you contact records. None of this is useful on its own.
The synthesis layer is where these fragments become coherent. It is the part of the agent that says "You have a meeting with Sarah in 30 minutes. She emailed you yesterday about the Q2 budget. The latest version of the budget spreadsheet is on your desktop. Here is the relevant section."
Why Synthesis Is the Hard Part
Calling APIs is solved. Every tool has an SDK. The hard part is connecting outputs across tools in a way that matches how humans actually think about their work. You do not think in terms of "email objects" and "calendar events." You think in terms of "preparing for my meeting with Sarah."
The synthesis layer has to:
- Resolve entities across systems - "Sarah" in email is the same as "Sarah Chen" in the CRM
- Understand temporal relationships - the email came before the meeting, so it is likely preparation material
- Prioritize by relevance - not everything about Sarah matters for this meeting
- Present a narrative - not a data dump, but a story that helps you act
Current Approaches Fall Short
Most agent frameworks skip synthesis entirely. They return tool outputs sequentially - here is the email, here is the calendar event, here is the file. The user has to do the synthesis in their head.
Some try to synthesize by feeding everything into a single LLM call. This works for simple cases but falls apart when the context is large or the connections are subtle.
What Good Synthesis Looks Like
The best synthesis is invisible. You ask the agent "prep me for my next meeting" and get a concise brief that connects all the dots. No raw data. No tool artifacts. Just the information you need, organized the way you think about it.
Fazm is an open source macOS AI agent. Open source on GitHub.