The Ann Arbor Architecture for Agent-Oriented Programming

TL;DR

The paper addresses the limitations of current prompt engineering and task-centric agent frameworks by reframing language models as automata and proposing the Ann Arbor Architecture as a unifying, agent-oriented paradigm. It introduces Postline as a prototype platform that uses an email-like MBox memory to support persistent, multi-agent collaboration, memory management via Memory Segment Rewrite, and dynamic agent creation across realm servers. The work provides concrete implementations and experiments, including a shell-robot workflow, code generation and execution, and binary data handling, to demonstrate the feasibility of continuous, episodic learning and memory-driven agent evolution. The proposed framework aims to enable more autonomous, contextually aware agents capable of operating across tools and environments, with potential impact in scientific and industrial research by leveraging episodic memory and retrieval-augmented capabilities.

Abstract

In this paper, we reexamine prompt engineering for large language models through the lens of automata theory. We argue that language models function as automata and, like all automata, should be programmed in the languages they accept, a unified collection of all natural and formal languages. Therefore, traditional software engineering practices--conditioned on the clear separation of programming languages and natural languages--must be rethought. We introduce the Ann Arbor Architecture, a conceptual framework for agent-oriented programming of language models, as a higher-level abstraction over raw token generation, and provide a new perspective on in-context learning. Based on this framework, we present the design of our agent platform Postline, and report on our initial experiments in agent training.

Figures (6)

• Figure 1: Definition of DFA by Hopcroft, Motwani and Ullmanhopcroft2006. With language models, a symbol is commonly called a token and the alphabet the vocabulary.
• Figure 2: Different paradigms of customer service agents. With existing platforms, development and deployment are separate processes with very different software environments. Under the proposed framework, both training and serving occur in conversations on the same platform and the transition is seamless.
• Figure 3: Different paradigms of research assistant agents. In the existing task-oriented paradigm, each tool has its own agent, and the researcher interacts with multiple agents that are unaware of each other. Under the proposed framework, a single agent operates across all tools to assist the researcher, maintaining continuous memory even when focus shifts between tools. Each tool implements the non-intelligent robot protocol so they can be automated by the agent.
• Figure 4: System architecture of Postline. The language models are invoked by the realm servers as external APIs and are not displayed in the diagram.
• Figure 5: Public-key encryption, not yet supported by language model APIs but anticipated, allows the model and the user to hide information from Postline and other intermediaries. The user pre-registers public keys on model servers to receive encrypted output.