Key points are not available for this paper at this time.
Language models (LMs) like GPT-4’s ability to predict word sequences have made tasks such as summarisation and translation significantly easier. They often struggle, however, with complex reasoning tasks that require deliberate, multi-step processes. To address these limitations, the concept of agentic artificial intelligence (agentic AI) is introduced, where LMs are organised into workflows that mimic human-like iterative reasoning. This paper explores the four pillars of agentic AI frameworks: tool use, reflection, planning and multi-agent collaboration (MAC). Tool use allows LMs to access external resources such as search engines to enhance accuracy; reflection enables self-correction and iterative feedback; planning ensures the structured breakdown of tasks; and MAC involves multiple LMs collaborating on specific subtasks. The paper then compares and contrasts various MAC frameworks based on the inclusion of tool use, reflection and planning abilities, as well as the flow and control of information between different agents. Finally, potential use cases for MAC frameworks are discussed.
Chawla et al. (2024) studied this question.