08 of 12
Agents
An agent is a model that does not stop at one reply. It chooses a next step, maybe uses a tool, looks at what happened, and continues until the job is done or it should quit.
In 2023, a wave of loops could browse, click, and write files. A refund, a deploy, or a coding task can fail on step twelve and still narrate success. Software engineering became the test bed because the world already has checks: tests, compilers, the list of changed lines.
Why do long jobs fall apart?
AutoGPT, in 2023, would loop on a to-do list and narrate progress until the bill stacked up. Each miss becomes the input for the next step. Even strong agents still fail a lot of real GitHub issues.
Why is code easier than booking a flight?
Software tickets come with tests. OpenAI and Anthropic shipped coding agents in 2025 and 2026 into that world. A desktop booking has no test that you chose the refundable fare, which is why the computer-use demos still stall on travel.
How does it know when to stop?
Someone gives it a budget of steps, a dollar cap, and a condition for quitting. AutoGPT did not have a good stopping rule. Codex and Claude in 2026 run in a hosted sandbox with a time limit so the loop cannot spend the weekend. Labs now collect traces of whole jobs, plus human judgments of which traces were actually good.