For two years the story of AI was the chat box: you ask, it answers. In 2026 the interesting systems have stopped waiting to be asked. They plan, call tools, check their own work, and keep going until a task is actually done. We call them agents, and they are quietly becoming the default shape of serious AI software.
From answers to actions
A language model on its own is a very good autocomplete. An agent wraps that model in a loop and a set of tools — a browser, a shell, an API, a database — and lets it take real steps in the world. The model decides what to do next; the loop lets it observe the result and adjust.
The tool-use breakthrough
What changed is reliability of tool use. Models now emit structured calls the runtime can execute and feed back. That single capability turns a text generator into something that can file a PR, book a flight, or reconcile a ledger.
pythonwhile not done and budget > 0:
thought = model.plan(state, tools)
result = tools.run(thought.action)
state = observe(state, result)
done = model.check(state, goal)Where it breaks
Autonomy multiplies mistakes as fast as it multiplies output. An agent that's right 95% of the time per step is right barely a third of the time over twenty steps. The engineering that matters now is not the model — it's the guardrails: scoped permissions, verification passes, and knowing when to stop and ask a human.
The model is the easy part. The loop around it is the product.
The teams winning with agents aren't the ones with the biggest model. They're the ones who treat the agent like any other unreliable distributed system — and design for failure accordingly.