Language Models Will Be Scaffolds
Source: https://alexzhang13.github.io/blog/2026/scaffold/ Author: Alex L. Zhang Date: 2026-02-26
Summary
Argues that the primary role of LLMs is shifting from “the thing that does work” to “the scaffold that orchestrates work.” The model becomes infrastructure rather than the intelligence endpoint. Current LLMs are underutilized because we treat them as monolithic answerers rather than compositional orchestrators.
Key Claims
- LLMs are already scaffolds in practice: they call tools, delegate sub-tasks, maintain state — they’re orchestration layers, not oracles.
- The inference-time compute shift: more useful to give a model 10x the turns than 10x the parameters.
- Scaffolds compose: an LLM scaffold can invoke another LLM scaffold, enabling recursive agent structures.
- The hard part isn’t the model — it’s designing the scaffold’s interface: when to delegate, what to remember, when to ask.
- Prediction: future “models” will be scaffold configurations (system prompts + tool sets) rather than model weights.
Connection to Other Sources
Complements Harness Engineering (encode discipline into environment). Aligns with Context Repositories (memory as scaffold component). Tensions with LLMs Not Getting Better — if merge rates are flat, is better scaffolding the answer?
Entities
- Alex L. Zhang — ML researcher/blogger
Concepts
- Coding Agents — primary example domain
- Context Engineering — scaffolds are context-engineering implementations
- Test-Time Compute — more turns > more params thesis