Feature Request: Sandbox Execution Isolation for Agent Runs
Feature Request: Sandbox Execution Isolation for Agent Runs: a task in LegoFlow-SWE (Harbor dataset). We need to add a sandbox profile to the evaluation harness so that different agent runs can be isolated from each other and from the host system. Currently, all agent tool executions run directly…
The task
We need to add a sandbox profile to the evaluation harness so that different agent runs can be isolated from each other and from the host system. Currently, all agent tool executions run directly on the host, which means a malicious or buggy agent could read protected benchmark artifacts, overwrite task documents, or…
Part of Lego-X/LegoFlow-SWE.