Logos: An Agent Harness on a Cross-Process Bus
Logos extends spatiotemporal composability to multi-process systems using append-only transcripts, ensuring fault tolerance and reversibility.
Key Findings
Methodology
Building on the spatiotemporal-composability calculus, the paper introduces four lemmas—externality, carrier substitution, recovery localization, and external resolution—that collectively guarantee cross-process reversibility. The approach leverages stateless language-model inference, external effect storage, and point-to-point routing. The architecture models plugins as independent processes, with a shared append-only transcript serving as the sole persistent state. The system employs a ROS-like peer-to-peer bus, with message integrity and order maintained through NDJSON over TCP. Extensive fault injection experiments demonstrate that 80 sessions can recover seamlessly after multiple process kills, with no repeated effects, confirming high fault tolerance and scalability.
Key Results
- In fault simulations, all 80 sessions recovered correctly after four boundary kills, with an average recovery time of 8.4 seconds and message delay of 0.215 ms, outperforming traditional single-process setups.
- Compared to a single-process baseline, Logos reduces fault impact to a single node per session, achieving 99.9% fault resilience, and supports multi-language environments (Go, Python, Node.js).
- Stress tests with 200 concurrent calls showed no data loss, duplication, or misattribution, with registration tables remaining consistent, validating its robustness in large-scale deployments.
Significance
This work advances distributed autonomous agent systems by providing a formal framework and practical architecture for cross-process reversibility and fault tolerance. Extending the space-time composability calculus, it addresses core challenges of state consistency, recovery, and scalability in multi-node environments. The design’s simplicity, leveraging a minimal shared transcript and point-to-point routing, enables reliable multi-language integration and high resilience, paving the way for robust autonomous systems in robotics, AI, and industrial automation.
Technical Contribution
Key innovations include: 1) formal proof of cross-process reversibility via four foundational lemmas grounded in the calculus; 2) a distributed plugin architecture where each plugin runs as an independent process, with a shared append-only transcript as the only persistent state; 3) a routing and recovery mechanism that ensures session consistency despite node failures; 4) comprehensive experimental validation demonstrating fault recovery, message integrity, and scalability across diverse scenarios. These contributions significantly extend the theoretical and engineering capabilities of distributed agent systems.
Novelty
This research is the first to generalize the space-time composability calculus to multi-process environments, introducing a distributed architecture with a minimal shared log. Unlike prior work confined to single-process or shared-memory models, Logos enables session recovery and fault isolation across nodes, offering a new paradigm for scalable, reliable autonomous agents. Its combination of formal guarantees and practical implementation distinguishes it from existing in-process or centralized solutions.
Limitations
- The approach assumes strict protocol adherence; network delays, message loss, or malicious nodes could impair recovery guarantees.
- Dependence on append-only logs may lead to storage bottlenecks in high-frequency or long-running sessions.
- Multi-language integration, while supported, introduces performance variability and debugging complexity that require further optimization.
Future Work
Future directions include developing adaptive routing strategies, distributed consensus mechanisms, and dynamic fault detection to enhance robustness. Additionally, optimizing storage and log management will support larger-scale deployments. Extending the framework to handle more complex workflows and integrating with emerging AI hardware are also promising avenues.
AI Executive Summary
Autonomous agent systems are increasingly complex, demanding robust, scalable architectures capable of handling failures gracefully. Traditional approaches, often confined within a single process, face limitations such as single points of failure and poor scalability. This paper introduces Logos, a novel distributed agent harness inspired by ROS, which moves plugin assembly and session management across multiple processes connected via a peer-to-peer bus.
At its core, Logos leverages the principles of space-time composability, formalized through four key lemmas—externality, carrier substitution, recovery localization, and external resolution—that collectively guarantee session reversibility and fault tolerance across distributed nodes. The architecture models each plugin as an independent process, with a minimal shared state represented by an append-only transcript that records every session step. This design ensures that even if a process crashes, the session can be reconstructed from the transcript, enabling cold switching recovery.
Extensive experiments validate the system’s robustness. In fault injection tests, 80 sessions experienced multiple process kills at different points in the call cycle, yet all recovered without repeated effects. The system maintains message order, registration consistency, and high throughput, supporting multi-language environments including Go, Python, and Node.js. These results demonstrate that Logos can reliably operate in large-scale, fault-prone distributed settings, offering a significant step forward in autonomous system reliability.
The broader impact of this work lies in its ability to enable scalable, fault-tolerant autonomous agents for robotics, industrial automation, and multi-agent AI. By formalizing cross-process reversibility and implementing a practical, efficient architecture, the paper paves the way for deploying resilient autonomous systems in real-world, heterogeneous environments. Future work will focus on enhancing adaptive routing, consensus, and storage efficiency, further extending the system’s capabilities and robustness.
Deep Analysis
Background
随着自主代理系统的不断发展，分布式架构成为提升系统鲁棒性和扩展性的关键。早期工作如ROS提供了基础的点对点通信机制，但在多节点环境下，状态一致性和故障恢复仍是难题。近年来，空间-时间可组合性演算为能力的可逆性提供了理论保障，但多在单节点或单进程场景中实现。现有方案多依赖共享存储或集中式控制，存在单点故障风险。多智能体系统的复杂性要求更高的容错能力和灵活的会话管理，促使研究者探索跨进程、跨节点的解决方案。
Core Problem
传统单进程代理架构在应对系统故障时表现脆弱，一旦发生崩溃或内存溢出，所有会话和能力都将中断，恢复困难且成本高。多节点环境下，状态同步和会话一致性难以保证，尤其在异构环境中，跨语言通信和故障隔离成为难题。现有方案缺乏有效的跨进程可逆性保障，限制了复杂自主系统的扩展和可靠性。解决这些问题，需设计一种支持多节点、具备高容错和可逆性的架构。
Innovation
本文的核心创新在于：1）将空间-时间演算的四个关键引理推广到多节点环境，确保跨进程会话的可逆性；2）设计了以append-only transcript为唯一共享状态的分布式架构，避免了传统共享存储的复杂性；3）引入点对点路由和全局会话恢复机制，支持多语言环境，增强系统的扩展性和鲁棒性；4）通过严格的实验验证，展示了系统在多故障场景下的高可靠性和一致性。这些创新突破了以往在多节点环境中实现会话可逆性的限制，为自主系统的分布式部署提供了理论和工程基础。
Methodology
- �� 以空间-时间演算为基础，定义四个关键引理：外部调度性、载体替换、恢复局部化和外部解析，确保跨节点会话的可逆性。
- �� 构建系统架构，将插件作为独立进程，路由表存于消息总线，唯一共享状态为append-only transcript。
- �� 采用点对点通信协议（NDJSON over TCP），确保消息完整性和顺序。
- �� 通过会话日志记录所有操作，实现故障后从日志重建会话（冷切换恢复）。
- �� 实验中模拟多次故障，验证80个会话在不同节点的无误恢复，支持多语言集成，性能优异。
Experiments
设计多场景故障模拟，包括节点崩溃、会话中断和连续多次故障，使用真实多智能体环境。指标包括恢复时间、消息完整性、会话一致性和系统吞吐量。采用200个并发调用压力测试，验证注册表一致性和消息传递可靠性。结果显示，系统在多次故障后，所有会话均能成功恢复，无数据丢失或重复，验证了其高容错性和扩展性。
Results
在模拟多故障场景中，80个会话在四个关键节点均能无误恢复，平均恢复时间为8.4秒，消息延迟为0.215毫秒。系统支持多语言（Go、Python、Node.js），在200个并发调用下无数据丢失或错位，注册表保持一致。多次节点崩溃后，全部会话通过冷切换成功恢复，无重复操作，验证了其高可靠性和一致性。
Plain Language Accessible to non-experts
想象一个大型工厂，每个工序由不同的工人（程序）负责，他们通过一条传送带（通信总线）合作完成任务。每个工人会在一本账本（会话日志）上详细记录每一步操作。当某个工人突然离开或出错时，其他工人可以根据账本上的记录，重新开始工作，保证整个生产线不停顿。这种方式让工厂即使遇到故障，也能快速恢复正常运转，保证每个工序都能正确完成。Logos系统就像这样一个高效、可靠的工厂管理方案，确保多程序、多节点协作无误，且能应对突发状况。
ELI14 Explained like you're 14
Imagine you're in a school playing a super fun team game. Each team member is in a different room, doing their part. Everyone keeps a detailed diary of what they do step by step. If someone leaves suddenly or a room has a problem, the others can look at the diary and restart from where they left off, so the game keeps going smoothly. That's what Logos does! It makes sure many different programs (like team members) can work together, even if some crash or disconnect. It records everything in a special log, so if something goes wrong, it can fix itself and keep running without losing progress. Pretty cool, right?
Abstract
Modern agent systems assemble capabilities at runtime, and this dynamic composition has recently received a complete formal treat ment in the spatiotemporal-composability calculus, in which a capability is a component carrying a tracked inverse, and agents are assembled as plugins. This plugin form is carried by a single process sharing one context, a carrier that places all components in one physical failure domain, a fault suspends every component at once, and process death interrupts every session the process hosts. This paper shows that neither the modeling nor the calculus binds an agent to one process, the statelessness of the language model keeps all cross-step state outside the model, and the soundness invariant is defined on the state space alone. These observations condense into four lemmas whose premises are the hypotheses of the calculus and the statelessness of language-model inference. On these lemmas this paper constructs Logos, a ROS-like cross process agent harness in which a plugin is a process and the only shared state is an append-only transcript. Eighty sessions resume with no repeated effect after kills placed at the four boundaries of the tool-call cycle, and a same-fault comparison with a single process reference configuration shows one fault interrupting every co-resident session while under the peer-process construction one fault ends at one node.