How it works
Each piece answers one question about how an agent behaves. You can set them up in any order.- Who answers. A custom agent has its own name,
@mention, instructions, reply language, and connections. - How it works. Skills hold your procedures and rules. Knowledge holds documents the agent looks things up in.
- What it can reach. Connections give access to your systems. Credentials hold other secrets, such as a vendor API key for a CLI. Use them when no connection exists for that tool.
- What it may change. Approvals decide whether each write runs, waits for a person, or is refused. There are two modes: Manual asks a person before every write, and Auto lets your Rules, then AI judgment, decide.
- Who hears about it. Notifications route alerts to email, Slack, Teams, or Google Chat.
/.
What’s per agent and what’s workspace-wide
A skill reaches every agent on each surface ticked under its Available in, such as Chat or Review. You can’t limit a custom agent to certain skills or knowledge bases. Scope it through its instructions and connections instead. Instructions steer the agent; they don’t block anything. Connections are the hard limit: an agent reaches only the connections attached to it.
Where each setting lives
In the sidebar, click More, then Workspace settings. The settings rail groups the pages like this:
For every permission by role, see roles.
What you can do
To choose between skills, knowledge bases, memory, Learnings and runbook sources, see Teach CloudThinker.
Custom agents at a glance
The number of custom agents each workspace can have depends on your plan:
The Create Custom Agent dialog shows how many slots you’ve used, for example 3 of 5 custom agents used. The organization Owner sees the plan itself in Admin Settings → Billing.
When you reach the limit, Create agent is disabled, and hovering it shows Custom agent limit reached. with an Upgrade plan link. Deleting an agent frees its slot right away. The limit is checked only when you create an agent, so after a downgrade your existing agents keep working, but you can’t add one until you’re under the new limit. For the other per-plan limits, see pricing.
Learning Mode
Learning Mode lets a custom agent turn a workflow it keeps repeating into a skill. It kicks in when the agent makes the same tool call on the same connection 3 times, in a conversation that has run at least 10 tool calls. You can’t change these thresholds in workspace settings.Only Admins can create skills in More → Workspace settings → Skills, but Developers can create agents. A custom agent on Auto or Offer adds workspace skills without an Admin, and they go live in Chat for every agent. Audit Logs don’t record them. Use Off for agents in production workspaces. No workspace setting enforces Off: an Admin sets it per agent in the agent’s settings, and can edit, re-scope, or delete a learned skill in More → Workspace settings → Skills.
Get started
Build your first custom agent
Create an agent, connect it, give it a runbook, and test it in 10 minutes
Skills
Turn your procedures into skills agents apply on their own
Approvals
Decide which writes run alone and which wait for a person
Agents
How CloudThinker, subagents, and custom agents share the work