Toward Systems Foundations for Agentic Exploration

2025-10-06

AI Summary

Existing snapshot-and-restore mechanisms (e.g., CRIU, container commit) struggle to support exploratory execution in LLM agents—such as branching, backtracking, and multi-path search—especially when external resources (files, sockets, cloud APIs) are shared, suffering from poor latency, low stability, and high overhead. This paper systematically identifies three core challenges: (1) branch semantics (visibility of cross-branch state updates), (2) external side effects (service-aware interception of API calls), and (3) native forking (microsecond-scale cloning of databases and runtimes). To address them, we propose a fork semantics model tailored for agent exploration, design a lightweight service interception framework, and implement a native fork mechanism that avoids full-state copying. Experiments show that off-the-shelf tools collapse under real-world workloads; our approach reduces branch latency to the microsecond level, enables high-fidelity shared-state exploration, and establishes a foundational runtime substrate for LLM agent systems.

Technology Category

• Multiagent Systems: Modeling other Agents
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Agent Architectures

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Agentic exploration, letting LLM-powered agents branch, backtrack, and search across many execution paths, demands systems support well beyond today's pass-at-k resets. Our benchmark of six snapshot/restore mechanisms shows that generic tools such as CRIU or container commits are not fast enough even in isolated testbeds, and they crumble entirely in real deployments where agents share files, sockets, and cloud APIs with other agents and human users. In this talk, we pinpoint three open fundamental challenges: fork semantics, which concerns how branches reveal or hide tentative updates; external side-effects, where fork awareness must be added to services or their calls intercepted; and native forking, which requires cloning databases and runtimes in microseconds without bulk copying.

Problem

• Defining fork semantics for agentic exploration updates
• Managing external side-effects in shared service environments
• Achieving microsecond native forking without bulk copying

Innovation

• Native forking clones databases and runtimes rapidly
• Fork semantics manage branch visibility of updates
• Intercepting external side-effects ensures fork-aware services