September 30, 2026
Agent-led code verification catches up with builds: Depot found a race condition in 21 minutes
In Depot's Jul 20 case study, a model checker explored 14,290,224 states in about 21 minutes and found a race condition that tests and review had missed. An agent translated the garbage collector redesign into TLA+, and verification proved 10 safety invariants and 2 liveness properties.
Gergely Orosz
@gergelyorosz
The last time software development sped up by roughly 10 times, in the early 2000s with the agile movement, interest in automated testing surged almost immediately: unit tests, TDD, XP and the rest. Now software development has easily sped up by 10 times too, and the same huge surge of interest in automated testing and verification is following...
· 7.1K views
In the early 2000s, faster development almost immediately spurred interest in unit tests, TDD and XP. Now AI agents have sped up software development by roughly 10 times, and interest in automated checking and verification is growing in their wake.
Tools are already available. Antithesis offers skills for AI agents through `npx skills add antithesishq/antithesis-skills`. Running them requires Snouty CLI, Docker or Podman, and npm. After `snouty login`, you can check a Docker Compose or Kubernetes configuration with `snouty validate <./path/to/config>` before starting a test. You will need to build the application for x86-64 and include dependencies that need internet access at startup in the image during the build.
Checking the model. Punt Labs' Z Spec plugin for Claude Code extracts specifications from code, generates code from specifications, checks types, and runs fuzzing and model checking through ProB. TLA+ makes sense for concurrent processes, complex transactions, autonomous workers and operations across systems without a shared transaction. A routine CRUD endpoint does not need a model checker. AWS has used formal specifications and model checking for critical systems since 2011: one bug with a trace spanning 35 high-level steps made it through design review, code review and testing.
The faster AI agents write code, the more often verification will start before a system launches rather than after a bug appears.
