OpenSpec puts a spec-review gate in front of 30+ coding agents
An MIT-licensed, propose-to-apply loop for your coding agent — while a fresh benchmark argues the harness, not the model, is what you're really choosing.
Copy markdown
One spec loop, 30-plus agents
OpenSpec (Fission AI, MIT) drops a propose → review → apply → archive loop in front of Claude Code, Cursor, Copilot, Codex and 30+ other agents — `npm i -g @fission-ai/openspec@latest`, then `openspec init` — and has cleared ~67k GitHub stars and today's HN front page. You sign off on the spec before the model writes a line, killing the 'it confidently built the wrong thing' tax.
The harness barely moves the score
HarnessTax benchmarked the scaffolding around coding models and found the harness barely changes accuracy — models often score higher in a rival's harness than their own, and a bare-bones 'Pi' harness matches Claude Code and Codex. Some system prompts hit 80KB, burning tokens for nothing, so your Claude models don't need Claude Code specifically and a leaner harness can cut cost without cutting quality.
Builders start trading their setups
A 'Show HN: Share your AI Setup' project hit 200 points today, letting builders publish and browse each other's toolchains, prompts and MCP configs. The through-line across all three stories: in late 2026 your setup — specs, harness, config — is the lever that moves output, not the model name.