CodeAF Launches as Open-Source Coding Harness for Open Models

CodeAF is a new open-source tool distributed as a single Go binary under the Apache 2.0 license, intended to maximize the utility of open models through specialized scaffolding. The team reports that CodeAF currently ranks first among popular coding agents on the DeepSWE benchmark, achieving the lowest cost per solved issue and outperforming mini-SWE-agent, which is the standard RL-training harness for many open models. The tool offers a unified interface for managing multiple coding tasks, splitting work into isolated repository copies, and intervening only when human judgment is required. Future developments include Pareto routing to optimize model selection based on cost and performance, expansion into specialist modules for security and finance, and a peer-to-peer architecture for cross-device continuity.