Trace Annotation

Module 1 · Lesson 1.3 · concept

Read a real agent trace the way a mechanic reads an engine: one turn at a time. For each turn, click every step you see happening in the trace. After you submit, each step turns green if you correctly identified it, or red if you mis-clicked a step that did not happen this turn. Missed-but-correct steps appear as a dashed outline so you can see what you overlooked.

The task

“Look at today's weather in Salt Lake City and recommend an outdoor activity for this afternoon. Save the recommendation to today-plan.md.”

The trace

Legend: correctly identified missed (actually happened) mis-clicked (did not happen)

Which termination type?

Termination Task complete (final answer, no tool call) Turn budget exhausted Unrecoverable error User stop

Rewrite the task

Rewrite the original task so a reader could look at the output and agree, without argument, that the task is done. One sentence.

What this teaches

Every agent-loop turn has five steps: prompt in, reason, decide, tool runs, result to state. Tool-call turns use all five; final-answer turns only use the first three. Naming each step on a real trace is the single highest-leverage diagnostic skill in this course. When an agent's output is wrong, the trace tells you which turn — and which step — failed.