Researchers define four conditions for LLM agents that survive process death without a host
An arXiv paper proposes that plugin-based agents can run across peer processes with no host, provided four conditions on state, records, recovery and providers hold.
In plain words
When an AI agent runs on its own, a crash of its main process usually wipes out its memory of the session. The paper describes four conditions that let an agent keep its state and records outside any single process, so it can recover after a fault. The authors also built an SDK, Logos, to test the idea.
What was announced
Large language model agents increasingly run long, multi-tool, autonomous sessions, and a single-process host keeps the session record, capability assembly and recovery basis in its own running state, so its death destroys them, the paper's abstract states. The authors present four conditions — on state placement, record carriers, recovery coordination and provider resolution — under which a plugin-based agent needs no host process. An agent meeting them can run its assembly on a bus of peer processes with its record in an append-only carrier owned by no process, and removing any one part makes a host necessary again. The conditions carry formal support from a composability calculus and supplementary proofs, the authors say.
To test the conditions, the authors built Logos, an SDK implementing them, and compared it against a single-process reference and a configuration over MCP, a protocol for connecting tools to models, with one fault injected in each of 200 tasks, realized in 197. The reference lost all 197 faulted sessions; the MCP configuration kept its sessions but issued 1,212 wasted calls; the SDK kept every session, wasted no call and lost no record, with task performance preserved. These are the authors' reported measurements, not independent results, and the abstract alone does not describe the tasks, workloads or how faults were injected.
What it does not show
- Evidence is limited to the paper's abstract; task design, fault-injection method and full experimental detail are not available for verification.
- The benchmark results are the authors' own measurements and have not been independently confirmed.
For security and IT teams
- Scope
- The paper addresses resilience of long-horizon, multi-tool LLM agent sessions against process death, not model quality or general agent capability.
- Status
- The work is an arXiv preprint (replace announcement) with formal support placed in supplementary materials; it is not described as peer-reviewed or as a released product beyond the described SDK.
Sources
Source-support review completed 14 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 14 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.