AI agents 2025 recap: What happened and what to expect next year

In 2025, AI agents, systems capable of using tools and acting autonomously, transitioned from research to practical application. This shift was fueled by advancements in large language models and the introduction of protocols enabling models to interact with external tools. Anthropic's Model Context Protocol in late 2024 was pivotal, setting the stage for the year of AI agents. Open-weight models and the Agent2Agent protocol from Google further accelerated the momentum, fostering communication between agents. These developments rapidly materialized in consumer products like "agentic browsers," which actively participate in tasks. Workflow builders also democratized agent system creation, lowering technical barriers. This new capability brought forth heightened risks, as agents' automation of tasks could be leveraged for malicious purposes. Future progress hinges on establishing agent benchmarks to evaluate both outcomes and processes, improving reliability. Governance, exemplified by the Agentic AI Foundation, will be crucial in setting shared standards for an open agent ecosystem. The optimal model size, whether large or specialized, remains a topic of debate as user control increases. Despite the optimism, significant socio-technical challenges persist, including infrastructure strains, job displacement, and intensified security vulnerabilities. Addressing these issues requires robust engineering practices, careful design, and clear documentation. A holistic approach, treating AI agents as socio-technical systems, is crucial to building a safe and innovative AI ecosystem.