AI foundations
A model, a workflow, an agent
Three ideas that get bundled together—and a practical way to decide how much autonomy a task needs.
Ask an AI system to summarise a document and ask it to investigate a broken application. Both requests involve language. They demand very different arrangements around the model.
A useful starting distinction is between the component producing an answer, the sequence organising the work, and the system deciding what to do next.
The model produces a response
Think of the model as one part of an application. It receives input and produces output: perhaps prose, a classification, or a proposed tool call. The surrounding software determines which documents reach it, which tools are available, and whether a suggested action is executed.
For example, a reading assistant might pass a short article and a request for five takeaways to a model. If that is the entire task, adding a planning loop brings little obvious benefit. First decide what a useful answer looks like: five distinct points, faithful to the article, with no invented details.
A workflow follows an arranged route
A workflow makes the sequence explicit. Consider a newsletter helper: gather approved links, draft short summaries, check that every summary includes a source, then present the draft for review. Different model calls may help with different steps, while application code controls their order.
Anthropic’s architecture guide uses a similar distinction: workflows follow predefined paths; agents let the model direct the process and tool use. The terminology varies across products, so inspect behaviour before trusting the label.
An agent chooses the next step
Now imagine a research task where the relevant documents are unknown. A system might search, inspect a result, notice a missing date, search again, and stop when it has sufficient evidence. That changing route is where an agent can be useful.
The extra freedom creates design questions. What counts as sufficient evidence? How many searches are reasonable? Which actions need review? What happens when two sources disagree? A fluent final answer cannot tell you whether these choices were sound; you need visibility into the work.
Choose with a small experiment
Take one real task and write down its input, expected result, and permitted actions. Try the simplest arrangement first. Keep several examples where it succeeds and several where it fails.
If the failures come from missing context, improve the context. If they come from a predictable sequence being skipped, make that sequence explicit. If they require different investigative routes depending on discoveries, consider an agent loop.
For a first agent experiment, use a bounded task such as producing a cited reading list from approved websites. Set a stopping condition and inspect its intermediate results. Expand its scope only after you understand the failures. The useful question is how much decision-making the task requires—and how you will check those decisions.