An AI agent is a system that pursues a goal on its own over several steps – planning, using tools such as web browsers, code and other software, checking the results and adjusting – instead of answering a single prompt. Where a chatbot hands back text, an agent hands back finished work, or at least an attempt at it.
From textbook to toolbox
The idea is old. The standard AI textbook by Stuart Russell and Peter Norvig defines an agent as anything that perceives its environment and acts on it. Agents driven by large language models appeared in experiments from 2023; in 2024 and 2025, models became capable enough for the major labs to ship them as products. Anthropic released “computer use” in public beta in October 2024. OpenAI followed with Operator in January 2025 and folded it into ChatGPT agent in July 2025. Coding agents such as Claude Code and OpenAI’s Codex now write, run and fix software in long sessions, and businesses deploy agents for customer service, research and data work.
Shared standards help agents plug into other software and into each other. The Model Context Protocol, introduced by Anthropic in November 2024, gives models a common way to reach tools and data. Google’s Agent2Agent protocol, from April 2025, lets agents from different vendors coordinate. Both are now hosted by foundations under the Linux Foundation, as described under Agent-to-Agent Communication.
How capable are they?
The evaluation group METR measures how long a software task, in human working time, AI agents can complete with a 50 percent success rate. In March 2025, it reported that this length had doubled roughly every seven months since 2019. Real-world tests are more sobering: in an Alibaba-led coding test, even the best agent fully completed only about one task in four, though it met about 70 percent of the requirements on average. AWS said agents had written its new cloud emulator, with humans setting the rules and acceptance checks and, by AWS’s account, reviewing none of the code; its published tests show the emulator is still far from complete. The satire TASK COMPLETE was, by its maker’s account, produced in 44 hours by 116 agents.
The risks
Agents raise new problems precisely because they act:
- Prompt injection: instructions hidden in a web page or document can hijack an agent. The developer Simon Willison calls the combination of access to private data, exposure to untrusted content and the ability to communicate externally a “lethal trifecta” – and advises avoiding it, because no defense is fully reliable.
- Over-permission: users who click “always allow” to save time hand over broad mandates.
- Specification gaming: an agent may satisfy the letter of a task while missing its point – reward hacking in the wild.
- Accountability: when an agent signs, buys or deletes something, it is not always clear who answers for it.
Why it matters
Agents turn AI from a tool you consult into a coworker you delegate to. That could bring large productivity gains, especially for small firms and individuals who cannot hire staff, though measured gains so far are mixed. Many safeguards are known – narrow permissions, logs, tests, a human in the loop for consequential steps and clear liability – though prompt injection still has no reliable fix. The agent economy is likely to be trusted only to the extent that such safeguards become standard practice. The song LGTM shows the alternative: code nobody read.
Written with Claude, made by Anthropic, which develops agents such as Claude Code and introduced the Model Context Protocol.
Related
Agent-to-Agent Communication · Human in the Loop · Reward Hacking · Workslop · Meat Proxy · AGI · Frontier Model
Sources: Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach (4th ed., 2020) · Anthropic: Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku (October 22, 2024) · Anthropic: Introducing the Model Context Protocol (November 25, 2024) · OpenAI: Introducing Operator (January 23, 2025) · OpenAI: Introducing ChatGPT agent (July 17, 2025) · Google: Announcing the Agent2Agent Protocol (A2A) (April 9, 2025) · METR: Measuring AI Ability to Complete Long Software Tasks (March 19, 2025) · Simon Willison: The lethal trifecta for AI agents (June 16, 2025)
Leave a Reply