An agent is a model wired to take actions, not just produce text. The mechanism is function calling: the model emits a structured request — 'search for X', 'run this code', 'query this database' — an external system executes it, and the result is fed back into the model's context. Loop that, and the model can plan, act, observe and adjust across many steps toward a goal. Standards such as the Model Context Protocol (MCP) provide a common socket so tools and data sources can be connected once and reused, rather than hand-wired into every application. The capability is genuinely new; so is the surface area for things to go wrong.

Byte: agents fail in compounding ways

A single model call has one chance to be wrong. An agent taking ten steps has ten — and errors compound: a wrong fact in step two poisons every step after it. The art of building agents is as much about constraining and checking each step as it is about capability.

1. 1.Function calling lets a model:

   Train itself on new dataEmit structured requests to external tools and use the resultsRun without any promptGuarantee correct answers

2. 2.The Model Context Protocol (MCP) is best described as:

   A benchmark for reasoningA type of GPUAn open standard for connecting models to tools and dataA pricing model