Planarian: Managing Agent State with Statepoints
- ID: 1db94e91
- 原文链接: https://arxiv.org/abs/2609.35366
- 作者: Jinnan Guo, Hao Mark Chen, Kapil Vaswani, Andrew Paverd, Peter Pietzuch
- 日期: 2026-09-28
- 更新: N/A
- 分类: agents
- 来源类型: paper
- 标签: agent-runtime, state-management, snapshot-rollback, fault-tolerance
- 质量评分: 4/5
- 抓取时间: 2026-10-01T15:57:56+00:00
中文导读
LLM agent 改文件、调本地工具、与远程服务交互，出错后的回滚目前要用户手动协调，本地与远程状态缺乏统一管理抽象。Planarian 给 agent runtime 加了 statepoint 抽象——环境状态的一致性可恢复时点，配三个原语：snapshot 用高效增量进程/文件系统快照捕获本地沙箱状态、并透明记录撤销远程变更的补偿动作（不要求外部服务支持 checkpoint）；rollback 回退本地检查点并重放补偿动作；fork 从一个 statepoint 分叉多个隔离分支并行探索。任务质量最高提升 15 倍，错误恢复开销仅 3%。
论文信息
- arXiv ID: 2609.35366
- 提交日期: 2026-09-28
- arXiv 分类: cs.OS, cs.AI, cs.CR
- 作者: Jinnan Guo, Hao Mark Chen, Kapil Vaswani, Andrew Paverd, Peter Pietzuch
- 链接: https://arxiv.org/abs/2609.35366
Abstract（arXiv 原文）
LLM agents solve complex tasks by iteratively changing files, invoking local tools, and interacting with remote services, which modifies state across their local environment and remote services. Today, agents and users must manage these changes explicitly, whether reverting exploratory actions or recovering from erroneous ones. Doing so safely requires coordinated actions, yet current agent harnesses lack unified abstractions and mechanisms for managing local and remote state consistently and efficiently. We describe Planarian, an agent runtime with state management that enables agents and users to recover from erroneous actions and explore alternative executions over consistent local and remote environment state. Planarian introduces the abstraction of agent statepoints, which are consistent, restorable point-in-time versions of the environment state. Planarian exposes three state-management primitives to agents and users: (i) snapshot creates a new statepoint spanning local and remote state without requiring external services to support checkpoints: it relies on efficient incremental process and file system snapshotting to capture local sandboxed state, and transparently records compensating actions to undo remote state changes; (ii) rollback restores the environment to a previous statepoint by reverting to a prior local checkpoint and replaying compensating actions for remote state changes; and (iii) fork creates multiple isolated branches from a statepoint, enabling the agent to explore alternatives in parallel. We show that Planarian enables agents to undo mistakes and explore alternatives in parallel, improving task quality by up to 15x, and allows users to recover from erroneous actions with only 3% overhead.
为什么值得关注
做 agent harness 的人值得把这三个原语与自己的检查点机制对照一遍——「补偿动作记录」的设计让不要求外部服务支持 checkpoint 成为可能。
English Summary
Planarian adds statepoints — consistent, restorable point-in-time versions of environment state — to agent runtimes, with three primitives: snapshot (incremental process/FS snapshots of local sandboxed state plus transparently recorded compensating actions for remote changes, no external checkpoint support required), rollback (restore a prior statepoint by reverting locally and replaying compensations), and fork (isolated parallel branches from one statepoint). Task quality improves up to 15x and users recover from erroneous actions with 3% overhead.