Benchmarks and reports from our work with AI coding agents.
Which model and effort level should do the coding work? DispatchBench runs Claude and Codex configs on fresh problems and repo tasks. It reports pass rate, cost and time per pass, and picks the cheapest config that still passes.
Claude Code on a personal server, reachable from a phone or ChatGPT, handing work to cheaper models and other agents. What it's made of, how it compares with Claude Duet, and what it costs.
Four articles from a six-week Ethereum client sync campaign: the write-up, raw results, agent methodology and harness reference.
A six-week sync campaign across Ethereum execution and consensus clients, measuring disk use, sync speed and restart resilience.
The raw results tables from the execution and consensus client sync campaign.
How a fleet of Claude agents ran, sampled and gated every sync during the six-week campaign.
A function-level reference for the harness used in the Ethereum client bake-off.