Planarian improves AI agent state management for better task recovery
Planarian: Managing Agent State with Statepoints
Operating SystemsArtificial IntelligenceCryptography and Security
Summary
LLM agents work by making changes across files, tools, and online services, which can get messy and hard to fix if errors happen. The authors created Planarian, a system that lets these agents save snapshots of their state, undo mistakes, and try out different options easily. Planarian keeps local and remote parts of the agent’s environment in sync and safe, which helps agents and people recover from problems faster. This approach improved task success by up to 15 times while adding little overhead.
What this means in practice
- •For software engineers: Manage and recover complex automated workflows involving local files and remote services with consistent state snapshots.
- •For devops teams: Enable safer experimentation and error recovery in automated service orchestration by isolating and restoring system states.
Authors
Jinnan Guo, Hao Mark Chen, Kapil Vaswani, Andrew Paverd, Peter Pietzuch
Abstract
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.