Living shim between humans and AI models
Canonical version: Living shim between humans and AI models.
Nate B. Jones (@natebjones) pointed at a software category that doesn't have a name yet: not a harness, more like a SHIM. A living, dynamic assistant layer that sits between us and increasingly powerful models, and manages the model on our behalf.
The core of the argument: this is different from a prompt, and different from an AI Agent Harness. A prompt is static; a harness wires a model into tools and loops. The shim's job is to compensate for a growing asymmetry: the models now have capabilities WE aren't fully aware of. Jones notes that even OpenAI seemed surprised by the power of its own unreleased models. When the tool outgrows the user's mental model of the tool, you need something in between that knows what the model can do and translates your intent into the right way of using it.
Put differently: Harness Engineering adapts the environment to the model. The living shim adapts the MODEL to the human. It would pick capabilities, calibrate autonomy, and shape interactions dynamically, so we get the benefit of frontier models without personally tracking every new capability drop.
I'd connect this to the supervision problem too. If AI makes it harder to remain idle because agents constantly wait for guidance, a capable shim is the missing manager in the middle: it absorbs the micro-decisions and only escalates what genuinely needs a human. Same direction as Agentic Context Management, one level up the stack.
Jones frames it as "where the puck is going" and expects startups to form around it.
References
Related
- AI Agent Harness
- Harness Engineering
- AI Agents
- Agentic Context Management
- AI makes it harder to remain idle
- Context Engineering
- Claudexor
About Sébastien
Ready to get to the next level?
Found this valuable? Share it with someone who needs it.