Isolation separates agent work by concern and limits each context to the information and authority needed for its objective. It reduces context bloat and authority confusion. section 7.5.1
Apply isolation through a few operational rules:
- Separate concerns such as research, coding, testing, and security; scope tools, files, memory, and runtime values to each task. section 7.5.1
- Keep instructions distinct from retrieved evidence and tool observations. Treat untrusted content as data to analyze, not instructions to obey. section 7.5.1 section 8.1.4
- Bridge contexts with structured returns containing findings, evidence, artifacts, risks, and unresolved issues—not raw logs or full exploratory content. section 7.5.1
Isolation adds coordination overhead. Use it when context limits, competing subtasks, or safety boundaries justify that cost; a single curated context can be simpler for small tasks. section 7.5.1














