Malleable software lets you reshape the tool yourself — adaptive software reshapes itself around you. Which kind of change can you actually see and trust?

This explores two ways software can change. In malleable software, the user does the reshaping. In adaptive UI, the system reshapes itself around the user. A caveat first: the collection has no papers that use either term, so this answer works from adjacent research. That research points to one useful idea. The real difference isn't who triggers the change. It's where the change lives, and whether the user can see and keep it.

Start with why this matters more now. Conventional software has a fixed context. Menus stay where you left them, so users can learn the interface and then reshape it on purpose. AI systems don't work that way. Their context (the prompt, the conversation history, the retrieved data, hidden state) shifts constantly and can't be learned the way a menu can How does AI context differ from conventional software context?. Put plainly, AI-driven interfaces are adaptive by default, whether or not anyone designed them to be. That changes what malleability means. It's less about editing a stable tool and more about getting a handle on something that's already moving.

The malleable side shows up in work on giving non-engineers direct control. Canvil lets designers shape a model's behavior through system prompts and structured trial and error inside Figma. The model becomes a material you can work with, not a black box Can designers shape LLM behavior without deep technical knowledge?. A warning sits next to it. When generative UI tools are given explicit design reasoning, they quietly fail to implement about a quarter of it, and about a third of the functional requirements Do generative UI tools actually implement their stated design rationales?. So malleability through natural language can be partly an illusion: you say what you want, and the system only partly does it.

The adaptive side shows up in work where the system learns from use. One line of work treats lightweight fine-tuning add-ons, called adapters, as persistent personal state. One shared base model plus millions of small per-user adapters stands in for millions of personalized models Can lightweight adapters replace millions of personalized models?. MetaClaw shows that this kind of adaptation runs on two clocks: fast skill fixes right after failures, and slower retraining during idle time Can agents adapt without pausing service to users?. A broader taxonomy sorts adaptation by two questions. Is the agent itself changing, or the tools around it? And is it learning from whether actions worked, or from how good the final output was How do agentic AI systems decompose into adaptation paradigms?? Neither question includes the user as an author. That's the quiet difference from malleable software.

There's also a surprising reversal. Sometimes the software changes and the AI has to adapt to it. Agent S handles constantly changing interfaces by combining web knowledge with layered memory of past tasks How can GUI agents adapt when software constantly changes?. Once AI agents operate software on our behalf, the old split between user-driven and system-driven change gains a third case: an agent adapting to software that is itself adapting. For a deeper treatment of malleable software and adaptive UI as design traditions, look outside this collection. It currently covers the AI mechanics underneath those traditions, not the traditions themselves.