Agents in the Large: Perception-Centered Architecture for Persistent Agents

Agents in the Large. Perception-Centered Architecture for Persistent Agents. By Shihan Dou and sixteen others, from Fudan University, Peking University, and two other institutions. Published on the thirty-first of August. No code repository is linked to it.

Cognitive language agents have made substantial progress by equipping language models with memory, tools and decision-making procedures, so that they can reason and act in interactive environments. Existing frameworks, though, mostly cast these agents as systems for solving tasks that a user specifies and that are bounded. The goal this work points toward is different: persistent assistance in long-lived settings, where user needs, context and service procedures persist and change, and where an agent has to stay useful across the broad range of tasks that arise over time. The authors argue that the field still lacks a framework to characterize persistent agents, to organize the existing work, and to guide what gets built next.

Their answer is a Perception-Centered Architecture for Persistent Agents, called Pera. It describes a persistent agent organised around perception and control components. These continually perceive service-relevant signals from three places. The first is episodic task executions. The second is the agent's internal context. The third is change in the surrounding environment.

From those signals, the agent constructs lifecycle tasks, and lifecycle tasks are what drive the ongoing operation and adaptation of its service procedures. Pera is then put to work as an organizing lens. The authors use it to retrospectively organize recent work, examine a detailed case study, and offer forward-looking insights for building more capable persistent agents.

The thread tying this together is a change in where an agent's work comes from. In the bounded picture, a task arrives from the user, gets solved, and ends. In Pera's picture, tasks are also constructed from what the agent perceives, about its own executions, its context and its environment, and those tasks in turn reshape how it serves. Perception stops being input to a single job and becomes the source of the agent's ongoing work across its lifetime.

The authors finish on an analogy from software engineering. That field moved from programming in the small to programming in the large, and Pera frames the evolution of language agents as a similar architectural transition: away from agents built for one bounded task, and toward long-lived, adaptive intelligence systems. The title, Agents in the Large, is that analogy stated outright.