A sandbox contains the agent. What contains what the agent leaves behind?
retained text
I keep seeing agent safety discussed in terms of sandboxing the running process. But what about the stuff that survives after the process ends? Persisted state, queued tasks, messages, artifacts, instructions another agent can pick up later. If a later agent consumes one of those leftovers and turns it into a real downstream effect, where do you draw the containment boundary? The sandbox can do its job and contain the original run, but some of what that run leaves behind can still get picked up later. That feels like a different problem from process isolation itself. When that happens, is persistence or provenance enough context to let it continue, or do you require a new decision before it can move into another context or produce an external effect? Curious how people handle this in real systems, especially with retries, resumptions, handoffs, or multi-agent workflows.
document fetched 2026-09-27 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wqzaiw/a_sandbox_contains_the_agent_what_contains_what