Skills & Tools
Settings → Agent is where you decide what each agent is allowed to do and what instructions it carries.
The page is organized by agent rather than as one flat list, because that's the thing that actually matters: the Better Sidebar agent can never use the file tools, and a list that mixed them together would imply otherwise.
Approval Policies
These three switches live in the Agent Dock, under What it may do. They apply to any running task.
Why reads default to on
Asking permission for every SELECT would make the agent unusable — a single question can involve a dozen queries. Reads can't damage anything, so they run.
Why writes default to off
Because they can. A misread instruction that renames forty conversations is a bad afternoon. The default is that you see each change before it happens.
Once you've watched it work a few times and trust its judgment on a particular kind of task, Don't ask again for this task in the approval prompt gets you a hands-off run without permanently loosening the setting. Turning Change data without asking on permanently is available, and reasonable if you rely on the agent daily — just know that undo only covers the most recent task.
Why "keep going on its own" is separate
It looks redundant next to the other two, and it isn't. The other two are about approval; this one is about whether the loop may advance unattended at all.
Turn it off and you review every round before it continues — without having to confirm every individual query. Some people want to watch the reasoning, not police the permissions.
The two are combined by intersection, which means this switch can only ever make things more manual, never less. Approving a write by hand never causes results to be sent onward behind your back.
Tool Groups
Each agent lists its tool servers with a switch, and expands to show the individual tools inside.
Better Sidebar agent
Better Sidebar data — query and change conversations, folders, tags, prompts and snippets.
Workspace agent
Files — read, write, edit and search text files.
Documents — read and edit Word, Excel, PDF and PowerPoint by outline and section.
Core
Always on for every agent, with no switch:
complete_task has no switch because a session that can't end isn't a preference anyone should be able to express.
Turning a tool group off removes those tools from the prompt entirely — the agent won't know they exist, rather than trying and failing. If you only ever want the agent as an analyst, switching Better Sidebar data off isn't the way to do it (that's the group it needs for reads too); leave Change data without asking off instead.
Skills
A skill is a block of specialized instructions the agent loads when a task calls for it. It's how the agent knows that spreadsheets should be read schema-first, or that tags should be reused rather than reinvented.
Built-in skills
Better Sidebar agent
Workspace agent
Built-in skills can be switched off, but not edited.
Writing your own
Click New skill on the agent you want it to belong to. Four fields:
Up to 20 custom skills.
The description does real work. It's how the agent decides whether this skill applies to what you asked, so describe the situation it's for, not just the action.
What a good skill looks like
Skills earn their keep on workflows you repeat. Instead of retyping a careful five-sentence prompt every Friday, encode it once.
Three things make this work rather than just being a long prompt:
- It stops for approval at a defined point, so you're not reviewing after the fact
- It constrains creation — folders and tags only get invented when genuinely needed
- It asks for a report, so you can tell what happened without reading every step card
Write the prompt by hand a few times first. Watch where the agent goes off the rails, add a sentence to stop that, and then save it as a skill. Skills written before you know the failure modes tend to be too vague to help.
Editing and deleting
Custom skills get pencil and trash icons. Deleting is permanent. Unsaved edits prompt before discarding.
Where the Agent Isn't Available
The agent needs an open chat with a working input box, because that's how it talks to the model. If you see "Open a chat first", navigate into any conversation and try again.
It also refuses to start a second task while one is running. Stop the running one first, or wait for it.