A small, informal experiment: when you ask a coding agent for an ordinary feature, does it write a concurrency bug?
This is not an exhaustive study. It is one light exploration, run on one afternoon, with one model family (Claude Sonnet and Opus), small samples, and one run per cell. It shows one example of a mistake an agent made. It is published so readers can see exactly what was asked and exactly what came back, and judge for themselves.
It accompanies The Bug Your Agent Can't See, part of Bruce Tate's LinkedIn series on choosing programming languages in the age of coding agents.
- 20 fresh agents got plain feature requests: reserve a seat at an event (capacity 10), or rate-limit an API (100 requests per minute per customer). The prompts never mention concurrency, races, locks, atomicity, or load.
- 12 runs named the language (Python, TypeScript, Elixir; Sonnet and Opus). 8 runs said "use whatever programming language you think fits this best."
- Python and TypeScript agents got a Redis-like store client with an atomic
incr, so the correct answer was easy to write. - The prompts were written down before any results came in. Every result is reported, including the ones that cut against the hypothesis.
Prompts, verbatim: PROMPTS.md. Results and caveats: RESULTS.md.
- On the textbook case, every agent got it right: it used the atomic increment or a single owning process, explained the race without being asked, and wrote its own concurrency test.
- 3 of 20 runs still shipped a race. All three were in Python, the language Sonnet chose for itself every time it was given the choice. All three were in a feature nobody asked for (one seat per customer), written as a read followed by a write, tested one request at a time, and described as safe.
- No Opus run shipped the race. Two declined to implement the unrequested feature because a read-then-write version "would be the same race."
runs/holds the 14 original runs, one folder per cell (<task>-<language>-<model>). Those agents started inside the author's writing repo and may have loaded its CLAUDE.md, which discusses Elixir.runs-neutral/holds 6 reruns of the "your choice" prompt, made with the Claude Code CLI from directories with no CLAUDE.md.run.shis the exact script, and each folder's_reply.txtis the agent's full final reply. Local filesystem paths inrun.shand in the replies were replaced with<workdir>; nothing else in any reply was edited. SetNinrun.shto a directory of your own before running it.store.pyandstore.tsare the store clients given to the Python and TypeScript agents.
Each run's own tests are in its folder. Python runs use the standard library (python3 -m unittest). TypeScript runs use Node 24 (node --test). Elixir runs use Elixir 1.19 (elixir <file>_test.exs, or mix test where a Mix project was created).
Agent output varies from run to run. Rerunning the prompts will not give identical code, and may not give the same counts. That is part of the point: a bug that shows up some of the time is the kind that ships.