May 12, 2026
Bots Behaving Badly? Not Anymore
Show HN: Statewright – Visual state machines that make AI agents reliable
AI coding babysitter drops, and the comments instantly split into hype, doubt, and receipts
TLDR: Statewright says it can make AI coding tools more reliable by forcing them to work in strict phases instead of letting them freestyle. Commenters were intrigued but split between “finally, useful guardrails” and “show us the proof, because this may just be extra complexity in a trench coat.”
A new Show HN launch called Statewright is pitching a very simple idea with very big energy: stop letting AI coding assistants run wild, and force them to follow a step-by-step workflow instead. In plain English, it puts the bot in a tiny fenced yard. First it can only read. Then it can edit. Then it can test. The creators say this makes smaller, cheaper models suddenly look much smarter, even claiming some local models went from repeated failures to perfect scores on a small software-fixing test set.
But the real show was in the comments, where the crowd immediately turned this into a mini civil war. One camp basically said, “Finally, some rules!” A few builders nodded along, saying they’ve made similar systems because today’s AI agents can be painfully slow, chaotic, and weirdly obsessed with rereading the same files forever. Another commenter gave it a polite but very loaded compliment: it looked like an existing state-machine company, just tuned for AI workflows — a classic Hacker News way of saying nice idea, but is it new?
Then came the skeptics. One blunt hot take argued that instead of adding a whole Rust-based rules engine, you could just have smarter AIs plan, cheaper AIs code, and other AIs review — basically, why build a traffic cop when you can hire more drivers? Another commenter zeroed in on the research claims and asked the dreaded internet question: where’s the code? That demand for receipts gave the thread its sharpest edge. The vibe overall: intrigued, slightly suspicious, and very ready to argue about whether “guardrails” are genius engineering or just extra plumbing with a fancy demo.
Key Points
- •Statewright uses deterministic state machines to restrict which tools an AI coding agent can access during each workflow phase.
- •The article positions state-based guardrails as an alternative to relying on larger models, longer prompts, or post-hoc observability.
- •Reported benchmark results show improved outcomes on a bug-fix task and a 5-task SWE-bench subset for several local models under Statewright constraints.
- •The system consists of a Rust engine that evaluates states, transitions, guards, and tool restrictions without an LLM in the loop, plus an MCP-based plugin layer.
- •The article demonstrates a bugfix workflow in Claude Code and describes guardrails including per-state tool enforcement, Bash restrictions, edit limits, and command allow-lists.