Reasonable Team Publishes TLA+ Tutorial: Not a Silver Bullet, AI Agents Could Change That
fhuszar · x · 2026-09-26
After Boris Cherny's viral tweet using Opus 5.5 to model the Claude Agent SDK in TLA+ and Lean (1M views), Ferenc Huszár's Reasonable team published a full tutorial. Key points:
- What TLA+ is: a 30+ year-old formal toolkit for describing allowed system behaviors and invariant/liveness properties.
- Not a silver bullet: TLA+ checks a model of software, not the software itself, and its main model checker only explores finite instances—so it can't fully verify implementations.
- The direction: combining temporal specifications, modern proof systems, and AI agents—from modeling behavior to generating machine-checked proofs—toward software specified, implemented, and verified in one loop.
- Their own work: a sneak peek at Reasonable's efforts to make this loop real, plus an interactive TLA+ playground.
The post also cites Datadog's harness-first agents as early evidence TLA+ pays off in agentic coding.
More from coding & agent
- Dev trains a 50MB model in 15 min to replace Gemini Flash at 0.06s latency — newz2000 · 2026-09-26
- DSPy 3.4.0 adds native System One model support with new ReAnchor optimizer — dair_ai · 2026-09-26
- Ex-MyFitnessPal CV engineer surveys industry: what's deployed, and what still hurts — kells1986 · 2026-09-26
- A CPU-Only Space Tags Every Hugging Face Release Into 48 Topics, No LLM — vanstriendaniel · 2026-09-26
- Opus 5.5-driven music-to-video pipeline: Blender to Loom via vgpu — DimitriDeJonghe · 2026-09-26
- Agent Harness notices user is absent and autonomously moves on without them — walkingriver · 2026-09-26