I don't think prompt-injection filtering can be a security boundary for agents on its own. OWASP maps nine dimensions of prompt injection, and a filter tuned for one misses the others, so enforcement has to sit at the action layer.
For each agent-to-tool integration, decide which layers you rely on: input filtering, a logged runtime check per action (allow, ask or block against a declared policy), and a sandbox that limits the damage if the model gets it wrong.
For an agent with database access, reads are allowed, writes need human approval, schema changes and exports are blocked, and everything runs in an isolated sandbox with scoped credentials. If you can only name the first layer, I think that is your real answer.