Coding agents need traces, gates, and spend controls before autonomy is credible
This week’s research treats large language model (LLM) agents as controlled software workers. The strongest work asks for traces, executable checks, tool limits, and review gates.