How to sandbox untrusted content in agent workflows
Treat external text as data, isolate tools and prevent instructions from crossing trust boundaries.
Start with the operational outcome
Reliable automation begins with a measurable result, defined inputs and an explicit failure path. The implementation should not receive broader authority simply because a model can perform more actions.
Use bounded stages
Separate discovery, planning, execution and verification. Give each stage only the tools and data it needs. Record inputs, outputs, costs and approvals so a failed run can be diagnosed without repeating irreversible work.
Verification before release
Test the normal path, malformed inputs, unavailable dependencies, duplicated events and partial completion. Customer-facing output should retain source traceability and be reviewed by deterministic checks before it leaves the system.
Where Oakhampton helps
Agent Desk turns this pattern into a scoped implementation with acceptance tests and operating documentation.