Nate B. Jones introduces a novel ‘living shim’, an adaptable intermediary layer designed to interpret intent, manage AI capabilities, and optimise human-AI partnerships, signalling a significant shift in how we collaborate with increasingly autonomous models.
Nate B. Jones has identified an emerging software layer that sits between people and increasingly capable AI models: not a prompt, and not quite a harness, but a living shim that can interpret intent, manage capability and decide how much autonomy the model should have. In David Sébastien’s write-up of the idea, the distinction matters because prompts are fixed and harnesses mainly wire models into tools and loops, while a shim would actively adapt the model to the human using it.
The case for that middle layer is rooted in a growing mismatch. As models improve, their abilities can outpace what users think they can do. Jones argues that even OpenAI appeared surprised by some of the power in unreleased systems, which suggests people will increasingly need something that understands the model’s real capabilities and translates them into useful action. In that sense, the shim would choose tools, tune autonomy and reshape the interaction on the fly.
That idea overlaps with broader work on human-AI collaboration. The Living Intelligence Project says it is exploring ways to move beyond simple prompt-and-response systems by building deeper coherence, initiative and shared reasoning between people and AI. Ethan Mollick’s co-intelligence framework, meanwhile, stresses that humans should stay involved, remain in the loop and treat each new model as the roughest version we are likely to use. Together, those views point to a future in which the quality of the partnership matters as much as raw model power.
There is also a more practical management question underneath it all. Jones’ notion of a shim aligns with the idea of reducing constant micro-decisions, so the AI handles routine choices and escalates only what truly needs a person. That is close to the emerging discussion around agentic context management and to recent arguments that organizational rules should be moved into code when teams become AI-native. Jones frames the concept as a place “where the puck is going”, and the surrounding discussion suggests startups are already beginning to think along those lines.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
