From LLM Reasoning to Autonomous AI Agents: A Comprehensive Review (2025)

This comprehensive survey covers the rapid evolution of AI agents from simple chatbots to autonomous systems that plan, act, and learn through continual interaction. It establishes a taxonomy for agentic AI and maps the field as of 2025.

The paper categorizes agentic AI into three tiers:

Foundational agentic reasoning: Core single-agent capabilities including planning, tool use, and search in stable environments. This is where tools like Claude Code and GitHub Copilot operate.

Self-evolving agentic reasoning: Agents that refine their capabilities through feedback, memory, and adaptation. These systems improve with use, learning from successes and failures.

Collective multi-agent reasoning: Collaborative settings where multiple AI agents coordinate, debate, and reach consensus. Papers like "From Debate to Equilibrium" propose frameworks for multi-LLM coordination.

2025 was the year AI agents went from research concept to production reality. This survey captures the field at an inflection point - the tools, frameworks, and techniques that make agents work reliably are maturing rapidly.

The key insight: agents are not just prompts with tools - they are long-running systems that need architecture, reliability engineering, and failure handling. The teams winning in 2026 treat agents as systems rather than clever prompts.

Real-world applications now span software engineering, scientific research, healthcare, finance, and education.

arXiv: 2504.19678. Scope: Covers AI agent frameworks from 2023 to 2025. Applications: Materials science, biomedical research, software engineering, chemical reasoning, and more.

arXiv: From LLM Reasoning to Autonomous AI Agents - https://arxiv.org/abs/2504.19678

arXiv: Agentic Reasoning for LLMs - https://arxiv.org/abs/2601.12538

GitHub: Awesome Agent Papers - https://github.com/luo-junyu/Awesome-Agent-Papers