StateFork: Branchable Infrastructure for Agent Exploration

2026-09-29

AI Summary

This study addresses the inefficiency and difficulty of branch recovery caused by external state modifications during multi-trajectory exploration in computer-use agents. We propose a logical control plane and checkpoint substrate architecture that decouples exploration strategies from physical states by constructing observationally equivalent yet physically efficient execution session abstractions. By integrating hierarchical file systems, process checkpointing, and persistent terminal-compatible command sessions, we implement the StateFork control plane and Waypoint substrate to enable efficient creation, restoration, and cleanup of terminal snapshots. Evaluated on Terminal-Bench under an equivalent node budget, our approach outperforms the Pass@20 baseline, achieving a 26% improvement in task completion accuracy and accelerating exploration speed by 70%.

Abstract

AI agents improve task success by exploring multiple trajectories, but for computer-use agents each trajectory modifies external environment state. Branching from an intermediate point is correct only when restoration is observation-equivalent - future actions produce the same observations - and practical only when creating, restoring, and discarding branch states is physically efficient. We study this problem for terminal-using agents, where tasks modify files, shell context, running processes, and local services. We introduce StateFork, a logical control plane that separates exploration policies from physical state materialization, exposing sessions, commands, snapshots, restores, and cleanup over multiple execution substrates. We also build Waypoint, a checkpoint/restore substrate for terminal execution sessions that combines filesystem layering, process checkpointing, and a persistent terminal-compatible command session. Together, StateFork and Waypoint improve terminal-agent exploration by combining sample-efficient search with efficient restoration of the right execution state. On Terminal-Bench, branch-based exploration through StateFork and Waypoint improves task completion over pass@20 at the same visited-node budget, and using Waypoint achieves 26% higher task accuracy than other execution substrates while completing exploration up to 70% faster. These results show that observation-equivalent, physically efficient execution sessions are a key systems abstraction for exploratory AI agents.

Problem

• AI agents
• state branching
• observation-equivalence
• terminal exploration
• checkpoint/restore

Innovation

• StateFork
• Waypoint
• checkpoint/restore
• observation-equivalent branching
• terminal-use agents