Agentic systems are typically constructed by granting an underlying model action-like capabilities through prompts, tools, memory, filesystem context, and execution routines. This paper describes a different design principle: instead of instrumenting a raw language model, one embeds an already capable coding agent as the inner execution layer of an outer, domain-specific agent scaffold. The result is a meta-agent whose performance gains arise less from larger model weights than from recursive composition, explicit artifact grammars, validators, and iterative runtime correction. We call this principle Meta-Agentic Bootstrapping. This is a conceptual position paper: it formalizes the pattern, distinguishes it from multi-agent systems and classical self-reflection methods, analyzes its mechanisms and failure modes, and proposes an evaluation protocol for future empirical work. No proprietary implementation details, private directory structures, or confidential scaffold contents are disclosed.
Geromichalos (Mon,) studied this question.