Ending agent approval fatigue: one isolated container per agent, then test the boundary
ComprehensiveShake76 · reddit · 2026-10-13
A Reddit practitioner shares three changes that made persistent agents actually usable:
- One computer per agent: each agent runs in its own Linux container — read-only system, only its own workspace mounted, no host credentials, dedicated browser and sign-in. With nothing personal inside, a task that used to need 3 approvals now needs 0. The author treats the container as one layer, not a guarantee.
- Test the boundary, don't trust it: the container sits on its own network; rules outside reject private and link-local ranges. The agent self-verifies on every start — EHOSTUNREACH for router/host, 200 from a public site — because VM restarts wipe firewall rules. Verified only on Colima/Apple Silicon so far.
- Make "done" checkable: research ends in a file with size, source count and hash (sources listed, not fact-checked); code ends with real test output, not the agent claiming it passed.
More from coding & agent
- Workday goes headless, opening HR and finance data to external AI agents — rseroter · 2026-10-13
- Decision model cuts main LLM calls 29.6% and cost 26.6% on 360 SOP-Bench tasks — vivekhaldar · 2026-10-13
- Your agent test can pass while it leaks your API key — a 200-line fix — rudra-sh · 2026-10-13
- clexo indexes Codex and Claude Code sessions into one searchable archive — preetam960 · 2026-10-13
- EMNLP paper: higher-accuracy LLM agents fail to flag uncertainty under knowledge conflict — DanielKhashabi · 2026-10-13
- 29 Bugs in Benchmark History Have Never Been Fixed by Any Model — PawelHuryn · 2026-10-13