Why AI agents need a two-tier architecture for tool execution and secret isolation
badhiyahai · reddit · 2026-07-24
The author argues that public-facing tool-using agents should use a two-tier design: one persistent server that only handles prompts and tool planning, and a separate ephemeral machine that actually executes code. This isolates malicious prompts so a bad command only affects one user’s sandbox instead of the shared server.
The post also notes a related secret-management pattern: never store OpenAI or other API keys on the persistent server. Instead, inject them on the fly through a proxy backed by a separate vault. The author says Anthropic’s managed agents follow a similar idea, even if they don’t call it that.
More from coding & agent
- ThoughtSpot launches SpotterCode, a hosted MCP server letting Claude Code query live instances — thoughtspotdev · 2026-10-26
- Non-developer stuck at 65% with Claude Code: rules files and CLAUDE.md weren't enough — jinguh · 2026-10-26
- HermesOBD: open-source macOS app pipes live car OBD-II data to a local Hermes agent — Teknium · 2026-10-26
- User Claims Codex Autonomously Reverse-Engineered Serum2 and Stripped Its License — Kyrannio · 2026-10-26
- colab-mcp lets AI agents control Google Colab notebooks directly — helloiamleonie · 2026-10-26
- Postiz, an open-source AI-agent social media manager, hits 37.9k GitHub stars — abhishek__AI · 2026-10-26