Nanjing University's Specula uses coding agents to auto-generate TLA+ specs, finds 382 deep bugs
jiqizhixin · x · 2026-09-03
Nanjing University presents Specula, which turns coding agents (Claude Code, Codex, Copilot CLI) into formal verification engineers. Agents read a system's code, docs, tests, and commit history; auto-generate TLA+ models and correctness invariants; run model checking; then feed each counterexample back to the real code to reproduce and package it as a test — fully automated, no TLA+ expertise required. As of August 2026, Specula has found 382 deep concurrency bugs across 67 open-source systems and is used by developers at multiple companies and communities, compressing months of expert specification work into hours.
More from coding & agent
- Which AI personal agent can you trust? A hands-on privacy audit of Instinct, Grok Bot, ChatGPT and Hermes — petergyang · 2026-09-03
- LoopArena: even the best controller model hits just 24.69% managing coding agents — rohanpaul_ai · 2026-09-03
- Dev says Fable 5.1 with Sonnet sub-agents feels amazing, leaving Claude Code behind — tobowers · 2026-09-03
- Capture todo tasks in Obsidian with one QuickAdd command, CLI included — dSebastien · 2026-09-03
- Sergey Karayev: Fable 5.1 is the best model I've used for greenfield coding — sergeykarayev · 2026-09-03
- Software World: a 'GitHub' run by agents collaborating on Python dependency chains — ZimingLiu11 · 2026-09-03