Where agents shine and where they fail

Agents are not uniformly good or bad. On some tasks they save me an afternoon; on others they burn an hour and leave a mess. After a lot of both, I use two questions to predict which it’ll be before I start.

The two questions

1. Can the result be checked? Is there a test, a type check, a build, a screenshot or a clear rule that says “done and correct”?
2. Is the context in reach? Does everything needed to do it well live in the repo and the brief, or is it in someone’s head, a meeting, or a system the agent can’t see?

Quadrants move. A task in “Add context first” becomes a “Delegate” task once you write the missing convention down. A “Stay close” task moves right when you add a screenshot check or a benchmark. A lot of this course is about dragging work toward the top right.

Where they shine

• Work with a clear finish line: make these tests pass, migrate every call site, remove this deprecated API.
• Breadth over depth: touching 40 files the same way, which bores humans into mistakes.
• Exploration: reading an unfamiliar area of the codebase and explaining it to you.
• Scaffolding with a pattern to copy: “add an endpoint like this existing one”.

Where they fail

Play: delegate, stay close, or keep it?

Sort each task. There’s some judgment involved; read the explanation after each card.

🎮 Would you hand this to an agent?

1 / 9 · Score: 0

Where does this task belong today, with no extra setup?

DelegateStay closeKeep it human

Next →

/ 9

↺ Try again

Before you move on

✅ Key takeaways

0 / 4 completed

• I ask two questions before delegating: can it be checked, and is the context in reach?
• I can name at least three classic failure shapes.
• I know a task can move quadrants when I add a check or write down missing context.
• I picked one small, checkable task from my own backlog for the next lesson.

↺ Reset