Hacker News front pageMay 12, 2026, 22:24 UTC+8

Show HN: Statewright – Visual State Machines for More Reliable AI Agents

Show HN: Statewright – Visual state machines that make AI agents reliable

Statewright uses a Rust state-machine engine to constrain Claude Code tool access, iterations, transitions, and guards; the post says 13–20B models improved consistently on real SWE-bench tasks, but it does not disclose benchmark scores, sample size, or the exact evaluation protocol.

Why it matters: HKR-H/K/R all pass: the state-machine constraint is a clear agent-reliability hook with a testable SWE-bench claim. Exact scores and reproduction details are not disclosed, so it stays just above the featured threshold.

Read the original ↗Export Markdown

Take

Statewright has the right instinct, but the evidence does not support the “make AI agents reliable” headline. It uses a Rust state-machine engine to constrain Claude Code tool access, iteration counts, transitions, and guards. That is real engineering, not another prompt wrapper. The weak spot is the eval: the post says 13–20B models improved consistently on real SWE-bench tasks, but gives no score, sample size, task split, or failure taxonomy.

I buy the mechanism more than the claim. A lot of coding-agent failure is not “the model cannot code”; it is uncontrolled shell use, file edits, loops, and bad recovery paths. State machines can cut those tail risks. But without SWE-bench numbers, Claude Code versioning, and a reproducible protocol, this reads like a good seatbelt for agents, not evidence that the driver got better.

Source: Hacker News front page · github.com
Published: May 12, 2026, 22:24 UTC+8 · May 12, 2026