On September 15, 2026, Notion released 3.7, themed “Agent Skills for Your Whole Team.” There’s a lot in this update, so I’m only covering the parts that matter for AI Agent workflows. Everything here is based on Notion’s official release notes.
The One Thing Worth Your Attention: Skills Can Leave Notion
The definition of Skills is plain: write how your team works as reusable instructions. The official examples include:
- Producing a monthly operations review in your team’s format
- Reviewing a feature proposal against existing standards
- Drafting a job description from a leveling framework
- Turning a draft into an executive-ready summary
- Turning messy notes into a PRD
You don’t need to write a config file to create one. Just chat with Notion Agent and it builds the Skill, and Notion will also suggest things from your work that could become Skills. The release notes don’t list a plan restriction for Skills, while several other features are explicitly limited to Business and Enterprise.
What really caught my eye is this line: a Skill can be downloaded as a SKILL.md file, plus approved supporting files, for Claude Code, Codex, Cursor, Gemini, or Grok. When you update it in Notion, the downloaded copy gets a “newer version available” flag.
In other words, Notion is treating SKILL.md as a cross-tool exchange format. If your team already uses Skills in Claude Code, you now have another route: Notion becomes the single source of truth for skills, non-engineers maintain them there, and engineers pull the same file into their local Agent. The difference is that Notion gives you the “newer version available” flag, which reminds you to update.
There are two ways to use them. One is typing / in the Notion Agent chat box, picking a Skill, adding context, and running it. The other is setting a Skill to run automatically. You can keep several related Skills in a Notion database, and when you share a single Skill or the whole database, everyone gets the latest version.
Custom MCP Connections: Agents Can Hook Into GitHub and Amplitude
You can ask Notion Agent to connect to tools like GitHub or Amplitude, and it walks you through the setup step by step. Once connected, the Agent can pull in context, check performance, or take actions inside those tools.
The official notes stress one design choice: any action inside a connected tool waits for your approval before it happens. This “confirm before writing” approach is the most practical safety valve right now for letting an Agent touch external systems. The feature is in beta and limited to Business and Enterprise plans.
Custom Agent: Sub-Agents, Post-Meeting Automation, Embedding in Docs
- Sub-agents: One Custom Agent can call other Custom Agents, each with its own instructions, context, permissions, and model. The main agent decides when to delegate and merges the results. The official case study is Remote, which reportedly runs an IT help desk with five Agents and saves $500,000 a year. That’s a customer claim quoted by the vendor, and I haven’t verified it independently.
- Post-meeting automation: You pick which meetings trigger a Custom Agent. Once the summary is done, the Agent gets the transcript and summary, writes progress updates, creates to-dos, or sends feedback into a database. Limited to Business and Enterprise.
- Embedding in docs: You can drop a Custom Agent straight into any document, and anyone on the page can use it.
- Agent SDK (public beta): Developers can continue a Custom Agent conversation from anywhere, trigger an Agent from their own app, and stream the results back. Limited to Business and Enterprise.
Models and Other Updates
- The model picker has been redesigned and now shows speed, intelligence, and cost indicators. The official list of newly available models includes Opus 5, GPT-5.6 Sol, and Kimi K3.
- The Insights API lets developers get a Custom Agent’s credit usage, completed run counts, and status, and also enable, disable, or cap an Agent. Limited to Business and Enterprise.
- There’s a new Developer sidebar and entry point for managing Workers, connections, personal access tokens, and run logs in one place.
- AI search is 50% faster, limited to Business and Enterprise.
- The inbox can ask an Agent to group notifications and flag what needs action.
- A new Notion Agents iOS app lets you capture ideas, voice notes, and photos on your phone, and assign tasks to Agents.
My Take
For a small team like mine, this release suggests three things:
- Skills are becoming a universal format. Since Notion can export a Skill as a SKILL.md for Claude Code, Codex, Cursor, and others, writing your team’s knowledge as Skills is less likely to lock you into a single tool.
- Management interfaces for multi-agent systems are starting to show up. Sub-agents plus the Insights API’s usage and cap controls mean the question isn’t just “can it run” but “what does it cost and what did it do.” That’s exactly where multi-agent systems most easily spiral out of control.
- Check your plan before getting excited. Custom MCP, post-meeting automation for sub-agents, the Agent SDK, the Insights API, and faster AI search are all limited to Business and Enterprise. The release notes don’t say whether other plans can use them. For Skills themselves, the release notes list no plan restriction.
The only source is Notion’s official release notes, so how well each feature actually works and how stable it is is something you’ll have to find out by trying it yourself.
