Sub-agent isolation limits prompt-injection risk by separating work involving untrusted external content from work involving sensitive internal data, using distinct permissions. section 4.8.8
Separate sessions or subgraphs so each workflow has only the access it needs:
- Give a research agent web access but no internal-system access; give an analysis agent internal read access but no outbound network access. section 4.8.8
- In LangGraph, assign different credentials to subgraphs; AWS Bedrock multi-agent orchestration offers per-agent IAM scoping. section 4.8.8
