Think about the sign-in book at the front desk. It doesn’t tell the full story of every meeting; it records who arrived, when, and sometimes why, so later, if something’s unclear, it gives you a trail. A system log does the same for software: it may record that a job started, a file changed, a request failed, or an email guard blocked a send. It’s raw evidence, not a polished summary, and boring until something breaks.
How it shows up
This shows up with AI agents. When Claude Code or Codex runs work, the session holds useful context: we treat session extraction like a meeting, where you worked through decisions, errors, and corrections, and logging it lets a future person or agent pick up the thread. Logs also matter when a hook blocks an email send or catches an unsafe command: if it only flashes a message, you have to remember what happened; if it logs the event, you can inspect it later.
Don’t confuse a log with a trace: a log records events, while a trace follows one request through multiple steps, like following one visitor through the building instead of reading the sign-in book. Good logs are also a source, tying an output back to what happened.
Why you care
For client work, “Claude did something weird” isn’t enough to debug: What did it run? What did it edit? What error came back? Without a log, everyone is guessing.