Perceive, Plan, Act, Self-Correct: An Architectural Framework for Goal-Directed Agentic AI Systems

Abstract

Large language model (LLM) agents that autonomously perceive their environment, formulate multi-step plans, execute actions via external tools, and self-correct based on feedback represent a paradigm shift from prompt-response AI to goal-directed AI. Yet the field lacks a unified architectural vocabulary: dozens of agent frameworks have emerged, each with bespoke abstractions, making principled comparison and reproducible evaluation difficult. This paper proposes the Perceive–Plan–Act–Self-Correct (PPAS) framework, a four-phase canonical loop grounded in classical agent theory (BDI, OODA, SOAR) and extended for LLM-era systems. We decompose modern agentic architectures into an 8-layer technology stack—from foundation models through orchestration, memory, tool integration, inter-agent protocols, planning, applications, to observability—and systematically map 15+ open-source frameworks to this stack. We validate the framework through three empirical studies: (i) a benchmark meta-analysis aggregating results from SWE-Bench Verified (n = 500), WebArena (n = 812), and GAIA (n = 466) covering 12 frontier agents, (ii) a comparative evaluation of five design patterns (ReAct, Reflection, Plan-and-Execute, Tree-of-Thought, Human-in-the-Loop) on standardized task suites, and (iii) an analysis of three inter-agent protocols (MCP, A2A, AG-UI) for multi-agent coordination efficiency. Results show that reflective planning with human-in-the-loop approval gates achieves the highest task-completion rates while reducing irreversible-action failures by 73%.