Back office ops · Production

How Notion Built Custom Agents: Rebuilding Four or Five Times to Reach an Agent-Native System of Record

The problem

Notion's early attempts to build agents starting in late 2022 failed repeatedly because there was no tool-calling standard, context windows were too short, frontier models were not reliable enough, and too much complexity was exposed to the model.

First attempt

Before a function-calling standard existed, Notion tried to fine-tune frontier models with a custom tool-calling framework and pursued partnerships with multiple frontier labs, but neither approach produced robust production-ready behavior.

Workflow diagram · grounded in source
1
Email triage trigger
trigger
“A live walkthrough of a Notion Custom Agent handling coworking space tenant applications by triaging email, enriching applicants with web search, and writing structured data into a Notion database”
2
Web search enrichment
ai_action
“enriching applicants with web search”
3
Structured database output
output
“writing structured data into a Notion database”
4
Manager agent routing
routing
“shared databases as primitives, agents invoking other agents, ‘manager agents’ supervising dozens of specialized agents”
5
Self-inspection feedback loop
feedback_loop
“agents that can configure themselves, inspect their own failures, and edit their own instructions — with guardrails around permissions”
Reported outcome

After being rebuilt four or five times, Notion Custom Agents launched as Notion's most successful product launch in terms of free trials and conversions.

Reported metrics
Launch success (free trials and conversions)most successful launch in terms of free trials and converting people
Agent rebuild iterations before productionfourth or fifth time
Meeting Notes growth impactone of Notion’s strongest growth loops
Reported stack
NotionMCPCLIGeminiGPT fourfireworks
Source
https://www.latent.space/p/notion
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After being rebuilt four or five times, Notion Custom Agents launched as Notion's most successful product launch in terms of free trials and conversions.

What tools did this team use?

Notion, MCP, CLI, Gemini, GPT four, fireworks.

What results were reported?

Launch success (free trials and conversions): most successful launch in terms of free trials and converting people; Agent rebuild iterations before production: fourth or fifth time; Meeting Notes growth impact: one of Notion’s strongest growth loops (source-reported, not independently verified).

What failed first in this deployment?

Before a function-calling standard existed, Notion tried to fine-tune frontier models with a custom tool-calling framework and pursued partnerships with multiple frontier labs, but neither approach produced robust pro…

How is this back office ops AI workflow structured?

Email triage trigger → Web search enrichment → Structured database output → Manager agent routing → Self-inspection feedback loop.