Agentic Software Engineering · Week 1

The internals of coding agents

This week

First session (lecture)
Build a minimal agent live, then read a real agent's requests: system prompt, tool list, loop and stop.
Second session
Studio: capture and read your own session.
Due
A1

Materials

Films

Two green tests

Made for the earlier edition of this course, which ran model-written code through a course launcher. The current course does not require one.

6:36 · subtitles: English, 中文 · download

A plausible patch that passes its two existing tests can still break the contract. A test derived from the contract (not from the patch) can expose the omitted rule. The same habit runs through the rest of the week: a generated program is not an executed one, context is not retraining, each event in an agent's trace establishes only its own thing, and a repaired helper can still leave its caller wrong.

Done is a claim

Made for the earlier edition of this course, which ran model-written code through a course launcher. The current course does not require one.

9:37 · subtitles: English, 中文 · download

An agent record that says read, write, then "done" with no further calls does not show that the bug is fixed. Report verified only when a passing check appears after the last write; the stop reason (final or step_limit) is reported separately, and a summary of the run is a claim too: it must keep the last failed check and be checked against the transcript.