We build Coder One, a coding agent harness with first-class support for System One via Jev from TypeSafe.

Watch on X: https://x.com/OpenAgentsInc/status/2102773483109335209

We benchmark Coder One against Opus 5.5 in Claude Code using TerminalBench v4.

The full run is still ongoing, but the early results are very promising (~45% cheaper for ~same success rate). We'll let Opus 5.5 summarize them in the picture below 👇 and full report here: https://t.co/39A03s4ziQ

All our harness+benchmark code are 100% open-source: https://t.co/IU2BJYcG5z

Hazarding a prediction if these results hold: All coding agent harnesses from the closed labs will no longer make sense to use as separate products, only constrained and orchestrated by a System One-centric harness like Coder One.

"Claude Code, we're here to eat your f*cking lunch."

Stay tuned for the full benchmark.

--

One more fun analysis from Opus 5.5 @ 2:33:00:

"- No single lab can be the router. A lab's harness routes to its own models. The best policy we found already mixes labs: Luna on some tasks, Opus on others. A router that works across labs is structurally something only a neutral party can build.

• The layer on top of the lab harness pays. Jev plus our opinionated front end put in front of Claude Code kept the pass rate (24/24) and cut cost 61% and time 37%. The biggest measured headroom is choosing a configuration per task: about 7 cents against $1.09 on the same tasks.
• Open, shared optimization is a real differentiator. Labs and closed router companies keep what they learn about routing and tuning. With NIP-OPT, an optimization is a signed, pinned record with its evidence: which configuration won, on which workload, measured how. Everyone can reuse and check it. If contributions accumulate, each operator's runs make every other operator's Coder better. That's a network effect no single lab can match, because no lab will share its competitors' results."

NETWORK EFFECT IS BACK, BABY

SEE YOU SOON