Fetching from the wire…
Public story · 2026-08-31 · high
Logos runs each AI agent as its own OS process and shares only an append-only transcript, not memory, so one crash can't spread to other sessions.
Why now: Teams already run several agent sessions on one machine, and the failure mode Logos tested for is the one that takes all of them down when a single agent crashes.
Researchers killed an AI agent's process at each of four points in a tool call, per a paper called Logos. All 80 test sessions resumed with no repeated effects, and one fault ended at one node instead of interrupting every co-resident session.
Most agent frameworks bundle capabilities as plugins that all run inside one process, so every agent session on a machine shares one failure domain. If that process dies mid-task, every co-resident session dies with it, whether or not it caused the crash.
Logos's argument starts from a narrower technical point. The model itself is stateless, and the state that carries over between steps already lives outside it anyway. Nothing about composing agent capabilities requires putting them in the same process. Logos instead gives each agent its own OS process, and the only thing those processes share is an append-only transcript, not memory, not a runtime.
The kills happened at all four boundaries of the tool-call cycle. Eighty sessions survived every one of them and didn't repeat anything on resume.
When a crashed subprocess takes an entire agent run down with it, that's an architecture choice, not a requirement.
Each link below shares sources, entities, or timing with this story.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (agent, architecture, boundary, session); pushes against this story (but).
Shared topic / Tension / Downstream implication
Overlapping topics (agent, already, boundary, session); pushes against this story (against); traces where this leads (implication).
Same source domain / Shared topic / Downstream implication
Reported by the same outlet (arxiv.org); overlapping topics (agent, architecture, boundary); traces where this leads (which means).
Same source domain / Shared topic
Reported by the same outlet (arxiv.org); overlapping topics (agent, argument, boundary, tool call).
Shared entity: Eighty / Shared topic / Earlier coverage
Both cover Eighty; overlapping topics (agent, already); earlier Eighty coverage from 2026-06-27.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (agent, capability); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (agent, already); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (agent, boundary); pushes against this story (versus).