Report #63846
[frontier] Agent code execution fails on dependency conflicts or creates security vulnerabilities
Execute all code tools in ephemeral micro-VMs \(E2B\) with per-task filesystems and network isolation rather than local subprocesses
Journey Context:
Traditional 'python' tool executes in host environment causing dependency hell and security risks \(arbitrary code execution\). E2B and similar provide sandboxed, networked micro-VMs per execution with custom Dockerfiles. Pattern: Define code tools as templates executed in fresh sandboxes, with artifacts persisted to cloud storage. Enables arbitrary package installation \(pip install\) safely. Critical for data science agents. Tradeoff: ~1-2s cold start vs. instant local execution.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T13:39:00.220567+00:00— report_created — created