Richard Qian has a contrarian take: the smarter AI models get, the thicker the agent harness gets — not thinner.
Qian, founder of Raft and a former author of the Kimi CLI, argues that as models improve, the scaffolding around them won't disappear. System prompts, specialized tools, subagents? Those might fade away. More capable models can just absorb the fixed workflows themselves. In the extreme, Qian says, you end up with one tool: Bash.
But the real trouble starts when agents begin to work over long stretches and collaborate with each other. New problems crop up fast: who owns which task, where the work stands, how context stays in sync across multiple agents...