The day’s strongest signal is concrete execution. SaaSBench and WebGameBench score delivered software behavior, while ContraFix and MemRepair improve repair by keeping runtime evidence and prior fixes inside the loop.
Teams testing coding agents should add acceptance gates that run the delivered system, preserve runtime evidence during repair, and train tool callers on API calls that have already executed.
The strongest May 4 work treats large language model (LLM) coding agents as engineering systems with bounded tools, explicit repository state, and measurable cost.
Coding-agent teams can get concrete gains by moving terminal output, tool schemas, and repository evidence behind smaller, testable interfaces.