The story
During 2025, AI systems began to do more than answer questions: they started to carry out multi-step tasks on their own. Given a goal such as 'fix this bug' or 'research this market', an AI agent plans the steps, uses tools such as a web browser, a code editor or a spreadsheet, checks the results and keeps going until it finishes or gets stuck. Software engineering saw the earliest and strongest adoption, as OpenAI, Google, Anthropic and startups such as Cursor all released coding agents, and agents that browse the web or operate a computer screen followed.
The idea is old, since AI textbooks have framed the field around 'agents' since the 1990s, but it became practical after ChatGPT in 2022 made large language models widely useful. A 2022 research paper called ReAct showed that language models could alternate between reasoning and taking actions, and early agent projects in 2023, such as AutoGPT, were exciting but unreliable, as the models lost track of their goal. Progress came from several directions: models trained with reinforcement learning to reason step by step, such as OpenAI's o1 in 2024 and the Chinese lab DeepSeek's R1 in early 2025; models trained to use tools and operate computers; and shared standards for connecting AI to outside software, such as the Model Context Protocol, which by late 2025 was supported by products from several rival companies and was placed under an industry foundation run by the Linux Foundation.
The loop is simple: the model reads the situation, decides on an action, the software executes it, and the result is fed back. The difficulty is reliability. An error early in a long task can compound, agents can be fooled by malicious instructions hidden in web pages (called prompt injection), and they can take actions that are hard to undo. A 2025 study by the research group METR estimated that the length of task that AI agents can complete had been doubling about every seven months, and researchers disagree about how fast that will continue and how much human oversight such systems will need. As of mid-2026, coding agents were in wide use, while Stanford's 2026 AI Index found that the best agents still scored below people on a standard test of everyday computer tasks.
Why it mattered
- Software development changed as programmers started to delegate whole tasks, such as writing, testing and fixing code, to agents.
- Shared standards such as the Model Context Protocol and AGENTS.md grew up so agents could connect to many tools, creating a new ecosystem of integrations.
- Questions about security, accountability and oversight became central, as agents can take real actions in the world.
- Businesses began to redesign workflows, and economists and policymakers started to ask how automating multi-step work would affect jobs.
Sources
- Measuring AI Ability to Complete Long Tasks METR
- ReAct: Synergizing Reasoning and Acting in Language Models arXiv
- Artificial Intelligence: A Modern Approach UC Berkeley
- OpenAI launches Codex, an AI coding agent, in ChatGPT TechCrunch
- Linux Foundation Announces the Formation of the Agentic AI Foundation Linux Foundation
- The 2026 AI Index Report: Technical Performance Stanford HAI
- DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning Nature (via PubMed Central)
Drafted by AI (Claude) from the sources above and checked against them by AI; no historian has reviewed this page. Spot a mistake? Suggest a correction. How pages are made
Community notes
Loading notes…