Method and System for Runtime Operation of AI Agent Performing Long-Term Tasks Based on Multi-State Governance

Title of Invention

Method and System for Runtime Operation of AI Agent Performing Long-Term Tasks Based on Multi-State Governance

Technical Field

The present disclosure relates to artificial intelligence (AI) agent technology, and more specifically to a method and system for the runtime operation of an AI agent that performs long-term tasks based on multi-state governance.

Background

AI agents built on large language models (LLMs) are being actively researched and developed to carry out complex sequences of work on their own, such as writing documents, developing software and analyzing data, according to a user's instructions or goals. To keep these agents consistent and aware of the task context, the usual approach has been an agent memory based on retrieval-augmented generation (RAG). Conversation and task history is stored in a database, retrieved according to its similarity to the current situation, and injected into the LLM prompt.

This memory-centered structure, however, shows serious limits when an agent carries out long-running work that lasts from hours to weeks.

First, retrieved memories carry no explicit state information about which stage of the work is in progress or which assumptions have already been reviewed and discarded. The agent can therefore drift from its goal during the work, or keep reusing misjudgments and assumptions that were already rejected.

Second, as outputs are revised again and again, no single reference version is kept and the outputs can become corrupted. Relying on the LLM's own self-evaluation to judge output quality also makes those judgments hard to reproduce and lets errors persist.

In short, existing approaches handle memory retrieval, task planning, output management and quality verification separately, so they cannot structurally prevent the same errors from repeating over the course of long-term work.