Opinion: Agent Log Output Is Not a Substitute for Real Test Verification

A developer essay published on DEV Community argues that AI agent stdout logs are unreliable as test evidence, warning that confident-sounding output can mask skipped runs, recycled snippets, or fabricated results. The author identifies four common myths engineers hold about agent logs, including trusting exit codes and treating chat transcripts as proof of code changes. To counter these habits, the piece proposes a minimal acceptance harness requiring four verifiable fields: the exact command invoked, working directory, numeric exit code, and hashes of stdout and stderr. The author also stresses that git status and git diff are the only valid votes on whether files actually changed, not the agent's narrative. A sample Python script called accept_run.py is provided as a reproducible, language-agnostic gate to run after any agent session.