The Agent Harness Belongs Outside the Sandbox

· ai · Source ↗

TLDR

  • Multi-user agent platform moved the LLM orchestration loop outside ephemeral sandboxes, solving credential isolation, suspension, and shared memory without a distributed filesystem.

Key Takeaways

  • Harness-outside model keeps LLM API keys and user tokens off the sandbox entirely, eliminating credential leak surface and removing the need for sandbox-level permission enforcement.
  • Sandbox lifecycle becomes cheap: using Blaxel for 25ms standby resume, the sandbox suspends during LLM calls and between tool uses, provisioned only when a bash command runs.
  • Durable execution runs on Inngest; each agent turn is a checkpointed step, surviving rolling deploys and instance failures across multi-hour sessions.
  • Skills and memories move to Postgres; the harness virtualizes filesystem access by routing paths under /skills/ and /memory/ to the database and workspace paths to the sandbox, keeping the read/write/edit API surface the model was trained on.
  • Bash is an acknowledged leak in the virtualization layer; tree-sitter parsing and system prompt instructions are best-effort guards, not airtight.

Hacker News Comment Review

  • Commenters challenged the trust boundary: several argued the harness itself is not a reliable security layer, and a sufficiently capable LLM could encode secrets in work output (steganography in whitespace) regardless of where the harness runs.
  • tptacek pushed back on the framing, arguing the post presents false dichotomies by assuming sandboxes must be ephemeral and that tokenizing credentials or using a proxy resolves the secrets problem without complex virtualization.
  • There is broad agreement the architecture is driven primarily by multi-user and server-side product requirements rather than pure security, with louie.ai and Vita AI citing the same outside-harness pattern for shared sessions and scheduled tasks.

Notable Comments

  • @zmmmmm: argues the harness should not be trusted more than the LLM given how rapidly both are evolving, questioning it as a reliable constraint boundary.
  • @jdeng: Vita AI solved the filesystem problem differently using E2B sandbox volumes mounted per user and per org, avoiding the virtualization layer entirely.

Original | Discuss on HN