Agent Stuck? Here's How to Get Unstuck

Stop first

When an agent is confused, piling on another instruction doesn't help. More text just makes an already messy context worse. Stop before you type anything else.

Diagnose before you repair

Don't trust what the agent says about its own work. Look at what it actually did: the diff, the logs, the files it actually touched. The agent's summary and the real change can be far apart.

Reset small instead of arguing

Arguing with a confused agent rarely helps — the bad context is still sitting there. A fresh, clean session with a tight, specific brief usually beats ten back-and-forth messages.

Git is your undo button

If the agent made a mess, go back to the last good state: git checkout or git revert (see Git & GitHub Basics). That's almost always faster than fixing the mistake inside the ongoing chaos.

Break the task smaller

Instead of "build the whole feature", give the agent one bite: one file, one function, one test. Smaller tasks are easier to check, and the agent strays less.

When to change course, when to just do it

If the agent gets the same error wrong twice, a third identical attempt rarely changes anything. Switch approach or switch model. And sometimes 10 minutes of typing it yourself beats 60 minutes of wrangling an agent — so just do it.