What Are Foundation Agents? A Deep Dive into Next‑Gen AI Architectures

In 2025 the hype around Agents keeps growing; the MCP protocol opens the Agent ecosystem and the A2A protocol fuels expectations for multi‑Agent environments. Most current Agents are simple extensions of large language models (LLMs) and still lack essential capabilities such as reasoning, long‑term memory, autonomous learning, and safe alignment for complex real‑world tasks.

To precisely define the gap to general intelligence and guide the next generation of Agents, MetaGPT and Mila together with 47 scholars from 20 top research institutions released a comprehensive review titled Advances and Challenges in Foundation Agents: From Brain‑Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems (arXiv:2504.01990).

The research involves contributions from MetaGPT, the Montréal & Mila AI Institute, Singapore Tech University, Argonne National Laboratory, University of Sydney, Penn State University, Microsoft Research Asia, University of Illinois Urbana‑Champaign, Hong Kong University of Science and Technology, University of Southern California, Yale University, Stanford University, University of Georgia, Ohio State University, King Abdullah University of Science and Technology, Duke University, Hong Kong Polytechnic University, Google DeepMind, CIFAR, and many others.

The authors introduce the new concept of a Foundation Agent , a technical blueprint rather than a specific implementation. A Foundation Agent is an intelligent system composed of modular components such as complex cognition, multi‑level memory, world models, reward/value mechanisms, emotion/motivation, multimodal perception, and action systems.

Part 1: Core Components – Building the Cognitive Foundation

A powerful Foundation Agent is a complex system of cooperating core modules, inspired by cognitive and neuroscience insights. The seven key components are:

Cognition Core : The "brain" that performs high‑level decision‑making, reasoning, and planning. Unlike current LLM‑driven agents, this core integrates logical, causal, and commonsense reasoning, hierarchical planning, uncertainty handling, meta‑cognition, and dynamic strategy adjustment.

Memory System : A multi‑layer memory architecture covering short‑term, long‑term, and working memory, including episodic, semantic, and procedural memories. It must support efficient retrieval, storage, forgetting, and generalisation while avoiding catastrophic forgetting.

World Model : A predictive model of the environment that captures physical laws, social norms, and other agents' behaviours, enabling prediction, planning, and counterfactual reasoning.

Reward and Value System : Mechanisms that evaluate actions and provide learning signals, supporting multi‑objective optimisation, intrinsic motivations (e.g., curiosity), and long‑term value estimation.

Emotion and Motivation Modeling : Simulated human‑like emotions and motivations that act as heuristics for rapid state assessment and strategy selection (e.g., fear triggers avoidance, curiosity drives exploration).

Perception System : Multimodal sensors that extract meaningful features from raw data (text, vision, audio, touch) and feed them to the cognition and memory modules.

Action System : The interface that turns cognitive decisions into executable operations such as natural‑language generation, code execution, robotic control, or virtual navigation, while considering feasibility, efficiency, and risk.

Part 2: Self‑Evolution – Towards Autonomous Intelligence

Beyond a solid cognitive architecture, a Foundation Agent must be able to self‑evolve: continuously learn, adapt, and improve without constant human intervention. The paper outlines four key mechanisms:

Optimization Space : Defining which components (cognitive strategies, memory contents, world‑model accuracy, perception abilities, action skills) can be optimized.

LLM as Optimizer : Large language models not only serve as cognition cores but also generate code, modify parameters, and propose new structures to improve other components.

Online and Offline Self‑Improvement : Online improvement occurs during real‑time interaction (e.g., reinforcement learning, world‑model updates), while offline improvement uses collected data for deeper analysis, architecture redesign, or large‑scale model iteration.

Self‑Evolution in Scientific Discovery : A self‑evolving Foundation Agent can autonomously hypothesise, design experiments, analyse data, and refine research strategies, accelerating scientific progress.

Part 3: Collaborative and Evolutionary Intelligent Systems – Building Collective Intelligence

When multiple Foundation Agents are combined, they form a Multi‑Agent System (MAS). The paper discusses MAS design, topology, collaboration paradigms, and evaluation.

MAS Design : Collaboration goals (individual, collective, or competitive) and collaboration protocols (rules and conventions) shape the system.

Topology : Static topologies (hierarchical, centralized, decentralized) vs. dynamic topologies that adapt based on feedback, implemented via search‑based, generative, or parameter‑based methods.

Collaboration Paradigms : Consensus‑driven, collaborative learning, iterative teaching, reinforcement, and task‑oriented interaction lead to discussions, debates, voting, and negotiation among agents.

Collective Intelligence & Emergence : Through continuous interaction, agents develop shared understanding and collective memory, giving rise to emergent behaviours such as trust, strategic deception, and self‑evolution.

MAS Evaluation : New benchmarks assess coordination efficiency, information‑transfer quality, and group decision performance, focusing on both task‑solving ability and adaptive interaction.

Part 4: Building Safe and Beneficial AI Agents – Alignment and Responsibility

As Foundation Agents become more capable, safety risks increase. The paper highlights three safety‑related aspects:

Security Threats and Countermeasures : Adversarial attacks, jailbreaks, goal drift, and unintended interactions require robust training, content filtering, formal verification, explainability, sandboxing, and permission limiting.

Alignment Problem : Ensuring that autonomous, self‑evolving agents remain aligned with human values and intentions, involving value learning, intent understanding, and ethical reasoning.

Future Directions : Developing reliable alignment techniques, establishing AI safety standards, and creating legal, regulatory, and societal frameworks for responsible AI deployment.

The authors conclude that while the path to general AI is challenging, the Foundation Agent framework provides a clear research agenda for building autonomous, collaborative, and safe intelligent systems.

Paper link: https://arxiv.org/abs/2504.01990

HuggingFace link: https://huggingface.co/papers/2504.01990

GitHub repository: https://github.com/FoundationAgents/awesome-foundation-agents