Inside the Agent Loop

Strip away the framework magic and every agent is the same machine: assemble context → call the model → parse the action → execute → append the observation → repeat until a stop condition fires. The state is the message list. That's it.

The deep part: the model sees its own past reasoning as ground truth. A wrong 'thought' at step 3 doesn't get corrected — it gets cited. The agent confidently builds on its own mistakes because they read exactly like its successes. This is why trajectory hygiene matters more than clever prompting.