Planarian: Managing Agent State with Statepoints
Planarian lets LLM agents snapshot, roll back, and fork local and remote state to undo mistakes.
Planarian is an LLM-agent runtime that keeps local and remote environment state consistent through restorable statepoints. Snapshot captures sandboxed process and filesystem state and records compensating actions for remote changes that lack native checkpoints. Rollback restores a prior statepoint, while fork creates isolated branches so an agent can explore alternatives in parallel. The authors report task-quality gains of up to 15x and user recovery from erroneous actions at about 3% overhead.
- Statepoints are restorable point-in-time versions of local and remote agent state.
- Primitives are snapshot, rollback, and fork for recovery and parallel exploration.
- Local capture uses incremental process and filesystem snapshots inside a sandbox.
- Remote undo records compensating actions without requiring service-side checkpoints.
- Authors report up to 15x task-quality gains and about 3% recovery overhead.
From the original article
1 min read at arxiv.orgLLM 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.
By Jinnan Guo · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.