Notion 3.7 turns team know-how into portable agent skills

Notion 3.7 gives teams a shared place to package repeatable work for AI agents. The September 15 release adds a skills library: reusable instructions that describe how a team handles tasks such as business reviews, product critiques, or onboarding checklists.

The practical change is portability. Notion says a managed skill can be downloaded as a SKILL.md package with supporting files, then used in external agent environments including Codex, Claude Code, Cursor, Gemini, and Grok. MustHave.ai’s independent analysis confirms the export workflow and adds an important qualification: the file carries the procedure, while models, tools, permissions, and output behavior still vary by destination.

Notion’s skills library stores process knowledge once

Skills target a common failure mode in team automation: the person who knows the process keeps it in a private prompt, chat, or document. In Notion 3.7, a team can store the instructions centrally, share them, and see when a downloaded copy is out of date.

That makes a skill closer to a versioned operating procedure than a magic prompt. A review rubric can travel with its approved reference files, but each receiving agent still needs its own access rules and acceptance tests.

Custom MCP connections add tools and an approval step

The release also expands Notion’s agent platform with custom MCP connections. Business and Enterprise users can connect external services, retrieve context, and allow agents to take actions. Notion describes confirmations before connected tools make changes.

That confirmation is useful at the interaction boundary, but it does not replace permission design. Teams should prefer read-only scopes where possible, isolate production credentials, and record which user approved a write. The release language and independent analysis identify the feature as a Business and Enterprise capability, so plan availability matters before deployment.

Sub-agents and the Agent SDK widen the workflow surface

Custom Agents can call specialist Custom Agents as sub-agents. A triage worker might classify an incoming report, while another agent gathers metrics and a primary agent assembles the response. Notion also describes an Agent SDK public beta for triggering conversations and streaming results from applications, plus Insights controls for usage and run status.

Those pieces move Notion beyond a document assistant toward an orchestration layer. They also create operational questions teams should answer early: who owns the final result, what stops repeated delegation, how are failed specialists retried, and which skill and model version produced each handoff?

A low-risk rollout starts with one skill

Choose a repetitive workflow where a person already reviews the output. Write success criteria, forbidden actions, and escalation rules into one skill. Export it to two target agents, run a fixed test set, and compare correctness, tool calls, latency, cost, and human corrections.

Do not assume identical behavior because the package is identical. Keep the Notion version, downloaded file hash, target-agent version, and granted scopes with the test record. That evidence will matter when a team updates the shared skill or adds a new connected tool.

Notion 3.7’s durable value is the combination of shared instructions, portable files, controlled connections, specialist delegation, and usage visibility. The next practical milestone is not a larger prompt library; it is proving that one versioned procedure remains reliable across the agent environments a team actually uses.

Sources

• Notion 3.7 release
• Independent Notion 3.7 analysis