Key Concepts
Least privilege: Give the agent only what it needs to do the job, and decide in advance where a human steps in. What the agent is allowed to do matters more than the agent itself. The same agent can be harmless or risky depending on its settings.
Allow self-updates: A setting that lets the agent modify its own Instructions. It is handy while you are improving the agent and risky once other people depend on it. When it is off, only a human can change the agent.
Quick Reference
The four guardrails
- What it can use: turn off anything the job does not need
- When it must ask: reading can run on its own, but anything that changes something outside Profound (saving a doc, sending a message) should wait for you
- Whether it can change itself: decide if it can edit its own Instructions
- Who can reach it: decide who can edit and who can only use it, and keep the list to people who need access
Set the guardrails in Profound
- Open Agent settings and review Capabilities: Browser, Web search, and Image generation. Turn off what the job does not need (a content brief does not need Image generation).
- Set tool permissions: reading tool on Allow, writing tool on Ask, anything the agent should never touch on Block
- Under Behavior, set When a tool needs approval to Ask
- Decide on Allow questions. On, the agent can pause to ask you for input. Off, it goes ahead with what it has.
- Turn Allow self-updates off once others depend on the agent
- Open the Share dialog, give Can edit only to people who need it, and set general access to Only people invited
Check that approvals work
- Run a test. The agent should work through its steps, pause at the writing tool, and continue after you approve.
- This needs a writing step on the agent, or there is nothing to approve.
Why it matters: the same agent, two outcomes
- Save only after you say yes: the worst case is you turn down a brief
- Save whenever it wants: the worst case is a weak brief sits in a doc your team trusts
Common failure to avoid
- Setting these up after the agent is already shared. Both an over-permissioned agent and one that anyone can change are easy to prevent beforehand and much harder to undo afterward.