How it works
The loop is simple. The model receives a goal plus a set of tool schemas. It emits a structured call, the runtime executes it, the result is appended to the conversation, and the model is invoked again. Stopping conditions cap the loop; approval gates hold destructive calls for a human. The ReAct pattern — interleaving reasoning traces with actions — is the ancestor of most current implementations.
Everything hard sits around that loop: tool descriptions the model can distinguish, results compact enough not to crowd the context, retries that do not duplicate side effects, and state that survives a crash mid-task.
Example
"Reconcile last month's invoices against the ledger and flag mismatches" is agentic: the number of steps depends on what the data looks like. "Extract the total from this invoice" is not — it is a single structured-output call. Teams often pay agent-level complexity and cost for work that is really the second kind.
Why it matters
Agents are where the industry's optimism and its measured results diverge most sharply. Per-step accuracy that looks excellent in isolation produces poor end-to-end completion once twenty steps must all succeed, and public benchmarks on realistic multi-step work still show large gaps. That arithmetic, not model quality alone, is what determines whether an agent deployment holds up.