How many tools should one agent have access to?

Start with the minimum toolset required to complete one bounded job. Give read access before write access, narrow scopes before broad ones and explicit escalation for irreversible actions. Every added tool should correspond to a tested capability, not a demo moment.

Read access first. It sounds obvious but we almost gave write access because the demo looked better that way. Glad we did not.

One tool per clear job at the start. If the agent keeps asking for another system, we look at the workflow before adding permissions.

I keep a simple permissions table with what it can read, draft, change and publish. The gaps become very obvious when you write it down.