Beyond a single call
A basic model answers from its own weights. An agent instead runs a loop: it reads the task, decides whether to call a tool, sees the result, then decides again. This repeats until the task is done.
The action loop
- Think about what is missing to answer.
- Act by emitting a structured tool call with arguments.
- Observe the returned result text.
- Repeat until enough information is gathered, then answer.
Each step feeds the prior observation back into the prompt, so later decisions depend on earlier results.
Why chaining matters
Many tasks need intermediate facts. To book a trip the agent might search flights, then check a calendar, then compute a price. No single tool answers the whole question, so the agent composes them.
Where it breaks
- The loop can run forever if no stop rule exists.
- A wrong early call poisons every later step.
- Long chains grow the prompt and the cost.
Good designs cap the number of steps and check progress between calls.
Key idea
Multi step tool use turns a model into an agent by looping think act observe over several tool calls, composing intermediate results into a final answer while a step cap keeps the loop safe.