How did chat models grow into agents that use tools?

A chat model can describe how to look something up, but it cannot actually look it up. It also struggles with tasks that smaller, simpler programs handle easily, like exact arithmetic or fetching a current fact.^[2] The fix was to let the model call external tools and use what comes back. That step is what turned chatbots into agents.

Two 2022 and 2023 papers mark the shift. ReAct showed that a model could interleave reasoning with actions: it thinks about a step, takes an action such as querying a Wikipedia interface, reads the result, and reasons again. Grounding its answers in real lookups cut down on confident wrong answers, and on two text-based decision-making benchmarks, ALFWorld and WebShop, it beat prior imitation-learning and reinforcement-learning methods by 34 and 10 percentage points of absolute success rate, using only one or two examples in the prompt to learn the pattern.^[1] Toolformer went a step further and trained a model to teach itself when to call tools, including a calculator, a question-answering system, two different search engines, a translation system, and a calendar, deciding which to use and what to pass to it. The training was self-supervised: the model proposed its own tool calls, kept only the ones that measurably improved its next-word predictions, and needed just a handful of hand-written examples per tool to get started.^[2]

Once models could use tools, the loose end was wiring. Each tool needed its own custom connector, so adding many tools to many apps meant a tangle of one-off integrations. In November 2024, Anthropic introduced the Model Context Protocol (MCP), an open standard for how a model and an outside tool or data source talk to each other, so a connector built once can be reused.^[3] That is the world we are in now: models that not only answer but act, through a shared way of plugging tools in.