The one thing · Engineering judgement
Your agent finished.
Can you explain the change?
A music app its author stopped understanding. A decoder checked against a reference. Two accounts make a useful case for keeping comprehension inside the work.
John Patrick Glattetre wanted better music recommendations in Coral, his self-hosted player. In an August account, he describes using Gemini and Claude to connect an audio-analysis library to C#. It worked, then ran out of memory after roughly 500 tracks. With Claude, he found a workaround: process tracks in separate, short-lived CLI processes.[1]
More features followed. When bugs appeared, he opened the code and found that he no longer understood much of his own project. The software had grown beyond its author’s working knowledge.
The anxiety is showing up elsewhere. On 17 October, designer @shedsgns described finding it harder to start from a blank canvas after working with Claude and Codex. That is a report about creative independence, with no test of coding ability behind it.[2]
Give the agent something to be wrong against
Another developer, VoxelBoy, describes a more concrete checking loop. Sound effects in his game glitched in a Rust Opus decoder. He compared all 64 effects with libopus, the reference implementation, then used focused tests to locate the first mismatch. Claude Code helped with fixes and regression tests.[3]
There is an inspectable result. PR #58 changes how the decoder handles a mode transition and adds a test comparing its output with libopus. The PR was still open in the 19 October capture. We inspected the code and test; we have not run them.[4]
The reference and test make the patch easier to challenge. Comprehension still has to be checked separately.
Decode the same input with libopus.
Locate the decoder behaviour that diverges.
Keep the expected behaviour checkable.
Review can still leave a knowledge gap
Review does not automatically resolve that problem. One Reddit developer describes small, atomic commits and several layers of review, then adds:[5]
This is a pseudonymous self-report without a linked project. It puts a practical question beside code review: can you explain how the changed part works?
Research gives that question some weight. Anthropic’s January experiment involved 52 mostly junior developers learning an unfamiliar Python library. The AI-assisted group scored lower on the immediate comprehension quiz. The study does not establish long-term skill loss in experienced users of coding agents.[6]
| Keep with the patch | What you should be able to say |
|---|---|
| The intended behaviour | What should happen for this input, and why? |
| A failure case | What breaks before the fix? What checks the result? |
| Your explanation | Which part changed, and where would you look if it failed again? |
Our take
For your next agent-written change, choose one behaviour you can verify independently. Keep the test or reference with the patch. Then explain the changed path in your own words before accepting it.
If you cannot explain it, ask for a walkthrough, simplify the implementation, or reduce the scope. This is our proposed working rule; these accounts do not prove it prevents skill loss.
The next bug may arrive after the agent’s context is gone. Leave yourself enough understanding to start.
Sources & context
- John Patrick Glattetre — Reflections on software engineering in the age of AI6 Aug 2026 · First-person account; background, not a new release. Read 18 Oct 2026.
- @shedsgns — working with Claude and Codex17 Oct 2026 · Subjective UI/design experience. Read 18 Oct 2026.
- VoxelBoy — the game’s audio decoder and libopus15 Oct 2026 · First-person account; the post also promotes the author’s game. Read 18 Oct 2026.
- opus-rs #58 — mode-switch redundancy fix15 Oct 2026 · Code and regression test inspected 18 Oct; header and table captured 19 Oct. Test not run by NO FEED. Read 18 Oct 2026.
- tb5841 — atomic commits, review and codebase knowledge17 Oct 2026 · Pseudonymous self-report; no project supplied. Read 18 Oct 2026.
- Anthropic — How AI assistance impacts the formation of coding skills29 Jan 2026 · Small randomized study of immediate learning, not long-term deskilling. Read 18 Oct 2026.