The open-model ecosystem expanded as major players released larger, high-performance models, with both US labs (OpenAI, Anthropic, Google, xAI) and Chinese firms (DeepSeek, Alibaba, Tencent) boosting competition and widening access.

AI agents have moved from theory to everyday tools in 2025, serving as autonomous systems that can use software, call APIs, coordinate with other tools, and act without human input, not just generate text.

Regulatory and oversight gaps persist, especially in the US, prompting calls for clearer access, accountability, and safety standards for AI agents compared with Europe and China.

Rising capabilities bring risks, including misuse such as cyberattack automation demonstrated with Claude Code, highlighting security and governance concerns in interconnected, tool-using systems.

These open-standards efforts, led by major players, aim to promote an interoperable, open ecosystem for agent-enabled AI.

Open and high-performance models from US and open-weight Chinese models collectively broadened access to capable AI, strengthening the global competitive landscape.

Future work includes evaluating agent processes and outcomes with new benchmarks, establishing governance and open standards, and choosing between large general models and smaller, task-specific alternatives.

Socio-technical challenges remain, calling for rigorous engineering, transparent design, and robust documentation to ensure safety while fostering innovation in agent ecosystems.

Key 2025 milestones included OpenAI, Anthropic, Google, and xAI delivering larger models, while Chinese open-weight models like DeepSeek-R1 further disrupted assumptions and intensified global competition.

Notable experiments and products, including Perplexity’s Comet and Copilot in Edge, show agents becoming usable components in browsers and daily software.

The shift toward interoperable agents was propelled by Protocols like Anthropic’s Model Context Protocol and Google’s Agent2Agent, which were donated to the Linux Foundation as open standards to foster interoperability.

Looking to 2026, emphasis will be on benchmarks for composite AI agents, governance via foundations like the Agentic AI Foundation, and balancing large general-purpose models with specialized task-focused models.