`Statewright`: visual state machines to harden coding agents

Reliability is improved by shrinking the agent's action space, not inflating prompts. Statewright enforces phase-specific tools and transitions, making smaller models more usable and frontier models less wasteful; worth watching, but missing hard benchmark numbers.

• —Each state limits tools, iteration count, and valid transitions, so planning, editing, and testing are separated by protocol rather than prompt discipline.
• —The reported sweet spot starts around 13B+ models: below that they can follow the machine but still lose enough context to make bad edits.
• —A Rust runtime handles guards and transitions without another LLM in the loop, while MCP plugins expose the workflow to tools like Claude Code.
• —The practical bet is cost control and fewer failure spirals: giving the model 5 tools instead of dozens reduces token waste and off-policy behavior.