firehose> #llmops

Agent-Mediated Software

When an agent — not a human — is the direct user of your software, the agent becomes a middle layer between the person and the product, and that layer is what finally makes software personal. The observation comes from someone who shipped both GUI web apps and skills and could compare them: the web apps drew a daily flood of complaints (couldn't find a button, didn't like a feature); the skill barely draws any. The reason is structural, not quality — a human isn't operating the skill directly, an agent is. If a skill's template defaults to something a user doesn't want (say, navigation dots), they don't file feedback and wait for the author; they just tell the agent to remove it, and it's gone. The output can therefore be different for every single person, so the one-size-fits-all compromise that generates GUI complaints dissolves.

Two consequences follow. First, distribution flips: instead of the user going to find the software, a capability wrapped as a skill comes to them inside the tool they already live in — no new app to install, no new interface to learn (the video predicts small productivity apps, mobile and web, get replaced by skills for this reason). Second, a design mandate: because a growing share of what we build will be consumed by agents rather than humans, product people must learn to design for agents. This is distinct from Agent-Computer Interface (ACI) (how an agent wields tools to act on the world); agent-mediated software is about the agent standing between a person and a product on that person's behalf.

Claims


Linked from