A large language model (LLM) learns from extensive text data to predict the next token—a word or word fragment—given preceding tokens. That ability lets it interpret varied requests and generate responses. Some models also produce thinking tokens before answering, allowing them to analyze a task or formulate a plan. section 1.2.1
In an AI agent, the LLM is the decision-making core, not the whole system. Tools provide ways to interact with the outside world, while an iterative loop lets the model choose tools and decide when to stop. A workflow differs: its execution steps are defined by the developer rather than chosen dynamically by the model. section 1.7
Choose an LLM-based approach—and a model—with the task in mind:
- Use one when inputs are varied or unstructured, such as natural-language requests or images. For predictable inputs and outputs, conventional logic may be faster, cheaper, and more reliable. section 1.4.1
- Weigh performance, cost, and development speed. Open models offer downloadable weights; closed models are accessed through APIs and offer features such as tool calling and structured output. section 2.1 section 2.1.1
From prediction to a research agent
Predicting “school” after “Because it’s summer vacation, I didn’t go to …” illustrates the basic training task. A research request—such as surveying recent work or extracting findings from PDFs—demands more: the LLM must interpret the request, but gathering information may require tools. section 1.2.1 section 2.0
Table 2.2 Required capabilities for unsolvable problems
| Missing capability |
Count |
|---|---|
| Web search and web page access |
11 |
| File reading |
6 |
This capability breakdown makes the tool-free limit concrete: among 20 GAIA problems, 11 needed web access and six needed file reading. section 2.4.3
Training data is not a dependable substitute for external information. In the same GAIA test, only five of 20 problems were judged truly solvable without tools, although some models answered more correctly, likely by recalling facts encountered during training. An agent cannot safely assume that recalled information is current or precise. section 2.4.3
