The Agent Loop: Think → Act → Check → Repeat

The four steps

An AI agent typically works in a loop:

1. Think: What's the next sensible step toward this goal?
2. Act: Call a tool (read a file, run code, hit an API)
3. Check: Look at the result — did it work? Any new information?
4. Repeat: Back to step 1, until the goal is reached or the loop is stopped

Why a loop instead of one step?

Because real tasks are rarely solved in one shot. A bug fix might need: read the file, understand the error, propose a change, run tests, check the result, adjust. Each round produces new information that shapes the next decision.

An example

Goal: "Fix the failing test." Round 1: the agent reads the error message. Round 2: it opens the affected file. Round 3: it changes a line. Round 4: it reruns the test. Round 5: test passes — done.

The danger: endless loops

Without limits, an agent can get stuck — retrying the same wrong fix forever, or racking up cost. That's why good agents need stop conditions: a maximum number of rounds, a budget cap, or a clear success condition.