•Contextual contamination problem
–Since AI agents are based on a single AI, their response speed is a strength, but they have a weakness that can cause context pollution if they are given too long and complex prompts. The more users try to improve the output quality, the more they face the dilemma of confusing the AI, and there is a limit to solving it with prompt engineering alone.
–
•Architectural Resolution
–The multi-agent system (MAS) solves the limitations of a single AI, and since only the necessary information is input to multiple AIs with specialized fields, it is a structure that can ensure overall quality. There are multiple types of MAS (such as those that pass tasks sequentially like bucket relays, and those in which multiple agents distribute and coordinate in real time), and they are used differently depending on the application.
–
•MAS has evolved into a "hierarchical type"
–Among MAS, the "hierarchical type" with sub-agents under the main agent (command tower) is a cutting-edge mechanism called "agent-type AI", which is expected to have a dramatic impact on labor productivity because humans are freed not only from the execution of tasks but also from the work of supervision.
–
•AI is effective
–While it has been difficult to measure the effectiveness of generative AI and AI assistant deployments in the past, AI agents have demonstrated their ability to replace human tasks. Agent-type AI, which bundles multiple AI agents, has been launched by major companies in various industries, including manufacturers and retailers, and quantitative effects have been reported one after another.
–
•15 PoCs (Proof of Concept)
–In this report, we carefully select 15 companies from various success stories that have shown quantitative effects, and explain business issues, measures, and implementation effects.