Dev builds AI-assisted concurrency testing tool that caught real bugs without TLA+ specs
cto_junior · x · 2026-09-23
In a thread, ctojunior shares how his team used AI to help build a concurrency bug testing tool. Unlike a TLA+ checker, which exhausts all execution orders up to some size, their tool only tries the orders the scheduler hands it — trading formal proof for no spec language to learn, yet still catching a real bug.
Key practices from the thread:
- If you suspect an AI-written model, first test it on a bug you already understand — it's easy to automate the same wrong assumption twice.
- Some failures were bugs in the model, not the code, so every failure is checked against real code and the model's own mistakes get saved as tests.
- The hypothesis library handles shrinking: a failure 40 steps into a random run gets reduced until only the steps that matter remain, then becomes a regular unit test.
- They split an undo operation into its real steps and made the lock visible to the scheduler — the tool then found the third failure the 'waiting for undo' ordering missed.
Related event: Engineers Use LLMs to Hunt Concurrency Bugs Tests Miss(4 posts)→
More from coding & agent
- Claude Code Orchestrator AO Hits 10K Developers, Founder Joins Solo Founders SF — julianweisser · 2026-09-24
- Hum-to-song app yue2-hum-to-song trends on Hugging Face — Mothersuperior · 2026-09-24
- One Prompt to GTM Agent: Polar Automates Lead Finding, Research and Follow-ups — EXM7777 · 2026-09-24
- Vibe-modeling real parts: Onshape MCP + Codex iterate a drip tray to V4 with humans only pressing print — hrishioa · 2026-09-24
- Kong, the 'agentic reverse engineer': LLM-guided decompilation recovers names, types, and call graphs in Ghidra — tom_doerr · 2026-09-24
- Yacine on RL for LLMs: everything behaves like training with a 400 gradient-step lag — cephaloform · 2026-09-24