Externalization Review of LLM Agents: The Source of Reliability Lies in Structures Outside the Model

• The reliability of LLM agents stems from externalization, which shifts cognitive load into explicit structures outside the model, rather than from changes to model weights.
• Protocols ensure interoperability and vendor independence by externalizing call syntax, lifecycle rules, permission boundaries, and exploration metadata.
• Harness engineering temporally coordinates memory, skills, and protocols, with OpenAI Codex and Claude Code converging on similar structures.
• Risks of error amplification between modules exist, and existing benchmarks underestimate the contribution of externalization, necessitating new evaluation axes such as transferability and recovery robustness.

Protocol and Harness Design

• Protocols such as MCP, A2A, and A2UI convert free-form reasoning into structured exchanges, providing runtime constraints.
• The harness forms a cognitive environment that manages the agent loop, sandbox, approval gates, observability, configuration, and context budget.

Evaluation and Limitations

• As an interpretive review without experiments, the basis for the claim of harness convergence is limited to two products: Codex and Claude Code.
• Because the contribution of externalization cannot be measured separately, there is a risk of misattributing it to model intelligence performance.