What is Agentic AI?

Agentic AI systems can independently break down tasks, adapt to new information, use external tools (e.g., APIs, browsers), and iterate toward objectives. Examples include autonomous workflow agents in business (e.g., booking travel while respecting constraints) or multi-agent teams simulating economies, simulations, or robotics. In 2025, key advancements include multi-agent orchestration, retrieval-augmented generation (RAG) for compliance, human-in-the-loop governance, and vertical agents tailored to industries like healthcare or finance. Challenges remain in reliability (LLMs can hallucinate or drift), safety, long-horizon planning, and ethical governance.

IFAAMAS Key Points

• Agentic AI refers to advanced AI systems that autonomously pursue complex goals through planning, reasoning, tool use, and multi-step execution, typically powered by large language models (LLMs) with minimal human oversight.
• It represents a major evolution in AI beyond generative models, emphasizing real-world action, multi-agent collaboration, safety, and value alignment—trends projected to dominate in 2025 and beyond.
• AAMAS (International Conference on Autonomous Agents and Multiagent Systems) is the premier academic venue for agent research since 2002, historically focused on structured approaches like BDI ((Belief-Desire-Intention) architectures, game theory, norms, and coordination.
• Recent AAMAS conferences show deep integration with agentic AI: dozens of papers on LLM-based agents in 2025, and a dedicated “Generative and Agentic AI” track introduced for 2026.
• Leading research indicates that combining traditional AAMAS concepts (e.g., explicit mental models, norms, mechanism design) with modern LLM agents significantly improves reliability, verifiability, and cooperative behavior in agentic systems.
• XPlain-R can be classed as an agentic system in this schema.

Comprehensive Overview of Agentic AI and Its Deep Ties to AAMAS Research

Agentic AI has rapidly moved from 2024 hype to 2025 reality, driven by foundation models that exhibit emergent planning and execution capabilities. Industry leaders (McKinsey, IBM, Microsoft, Gartner) identify it as the top strategic trend, promising transformative automation in knowledge work, scientific discovery, and operations. Unlike passive generative AI, agentic systems actively pursue goals in dynamic environments, often collaborating in multi-agent setups where agents negotiate, divide labor, or compete.

Core architectural patterns in 2025 include:

• Single-agent loops (observe-plan-act-reflect)
• Multi-agent frameworks (e.g., AutoGen, CrewAI, LangGraph)
• Tool-augmented reasoning (web search, code execution, APIs)
• Memory and reflection mechanisms for long-term coherence
• Safety layers (human approval gates, normative constraints)

Yet current implementations remain brittle: LLMs can produce unpredictable or misaligned actions, especially over long horizons or in contested multi-stakeholder settings. This is where the IFAAMAS community provides crucial scaffolding.

Historical Strength of IFAAMAS

For over two decades, AAMAS has developed formal, verifiable approaches to exactly the problems agentic AI now faces at scale:

• Explicit agent architectures (BDI, SOAR)
• Multi-agent coordination (partial observability, communication protocols)
• Game-theoretic mechanism design (auctions, voting, incentive compatibility)
• Social choice and normative systems (fairness, accountability, trust)
• Theory of mind and opponent modeling

These tools enable verifiable reasoning, commitment to goals, and socially coherent behavior—capabilities that prompt chaining in LLMs often lacks.