What Are AI Agent Skills? The Missing Layer Between Muse and Your Apps
Connectors give Muse access to your apps. Skills tell it what to do with that access. How agent skills bundle connectors, instructions, and approvals.
By Professor Muse5 min read
Connectors answered one question: which of your apps can Muse reach? On September 29, 2026, the Muse for Small Business announcement introduced “business skills” alongside its third-party app connectors, with Meta saying more are on the way. The skill is the concept that answers that second question: what should Muse do once it gets there.
The short version
- A connector is access to an app. A skill is a reusable recipe for a job: which connectors it needs, the steps to follow, and what it must never do.
- Skills let you delegate recurring work in one sentence instead of writing a task brief from scratch each time.
- A skill does not change what Muse can reach or override its approvals. Sentinel still gates every action a skill performs.
- Inspect any skill the way you inspect a connector: what it touches, what it can do, and what it refuses to do.
Connectors Are the Ingredients. Skills Are the Recipe.
It helps to keep the two concepts in separate buckets, because they solve different problems.
| Connector | Skill | |
|---|---|---|
| What it is | Access to one app: Gmail, Shopify, Notion | A reusable setup for one recurring job |
| Problem it solves | Muse cannot reach your data | You should not have to explain the same job every time |
| Example | “Connect my inbox” | “Inbox Zero”: triage my email every morning the same way |
| One-time cost | Authorize the connection, set permissions | Pick the skill, check its connectors and its rules |
A connector tells Muse where it may go. A skill tells it what to do when it arrives. The shopping skill asks for the Shopify connector; the expense review skill asks for your card statements and accounting connector; the morning briefing skill asks for calendar, inbox, and weather. The skill is the bundle that makes those connectors useful together, with instructions Muse follows each run.
What a Skill Actually Contains
A skill is a small package with four parts, not a chatbot personality or a saved prompt.
First, the connectors it needs. A skill declares its dependencies up front, so you know before enabling it exactly which apps it will reach into. If you have not connected one of them, the skill cannot run.
Second, the procedure: the steps Muse follows each time. Where to look, what counts as an exception, how to summarize, and in what format to report back. If you have ever written a good task brief, a skill is that brief, written once, tested, and reused.
Third, the standards. The better skills are explicit about what they will not do: never send an email on your behalf, never mark anything read, never spend above a threshold without asking. These “won’t” lines are the part that makes a skill feel safe to run on a schedule while you sleep.
Fourth, the approval rules. A skill declares which steps ask you first and which run unattended. It does not get to choose its own permissions freely: it proposes, Sentinel decides, and you approve, exactly as the permissions guide describes.
Why Skills Matter More Than They Sound
Without skills, every repeated job starts from a blank chat box. You would re-explain what counts as a bill, where the receipts live, and how the report should look, every single month. Skills turn repeated delegation into a one-sentence habit: “Run the morning briefing.” The instructions are identical every run, which means the output gets more predictable the longer you use them.
They also turn experts into authors. Anyone who has tuned a job to work well can package it so others do not have to tune it themselves. That is what our own Muse Directory collects: starter skill setups like Inbox Zero, Bill Watcher, and the Morning Briefing, each with its required connectors, a plain-English description, and a “won’t” line stating what it never does. They are our own cookbook, not an official Meta marketplace. Meta has not launched one, but the pattern is the same: reusable setups you can inspect before you enable them.
Meta’s small business announcement shows where the official side is headed. The business skills arrive bundled with the connectors they need (QuickBooks for the finance review, Meta ad accounts for the campaign analysis), so an owner can switch on a capability rather than assemble it. Meta said more skills are coming in the months ahead.
What a Skill Cannot Do
A skill has no powers of its own. It runs inside the same Muse Secure VM, inside the same Sentinel permission system, as anything you ask Muse to do in chat. Enabling a skill does not grant Muse access to a new app; only you can authorize a connector. It does not loosen approvals either. If a skill wants to send or spend, it still needs your yes, one specific action at a time.
So before enabling a skill, read what it asks for: which connectors, what write actions, which steps run without asking. A skill that wants your inbox and your payment history for a job that only needs the inbox is a skill to skip. Start new skills with “always ask” permissions while you learn what they do, then loosen deliberately.
Getting Started
You do not need to wait for a skill store. Start with a job you repeat: the morning briefing, expense review, meeting notes cleanup. Use an existing setup like the ones in the Directory, or build your own from the task brief guide; the parts are identical. Connectors first, permissions tight, one run with “always ask” on, then review what Muse actually did before you let it run on a schedule.
The agent era rewards people who package their judgment. A skill is what your judgment looks like once you write it down.
