← sources

What ended approval fatigue for my agents: give each one its own computer, then test what it can't reach

source: reddit (reddit:AI_Agents)published 2026-10-13 13:39ingested 2026-10-14 02:36

retained text

Three changes made persistent agents actually usable for me. Sharing in case it saves someone a few weeks.

1. One computer per agent. Each agent runs in its own Linux container on my Mac: no capabilities, a read only system, only that agent's workspace mounted, no home folder or host credentials, and its own browser and sign in. With nothing of mine inside, I stopped approving every command. A task that used to ask me 3 times now asks 0. A container is a strong boundary, not a guarantee, so I treat it as one layer, not the whole answer.

2. Test the boundary instead of trusting it. The container sits on its own network, and rules outside it reject private and link local ranges. When the agent itself tried, it got EHOSTUNREACH for my router and my host, and a 200 from a public site. It re-checks that on every start, because a VM restart wipes the firewall rules. So far I've only tested this with Colima on an Apple Silicon Mac, not Docker Desktop, OrbStack or a Windows PC.

3. Make "done" something you can check. Research ends in a file with a size, a source count and a hash, plus a note that the sources are listed, not fact checked. Code ends with the real test output, not the agent saying it passed.

What do you check before you leave an agent running on its own?

document fetched 2026-10-14 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x4wwhu/what_ended_approval_fatigue_for_my_agents_give