Agentic AI
Study maps the inner workings of 11 production coding agents, finds no standard frameworks
An academic audit of leading AI coding assistants reveals they all build custom runtime systems instead of adopting shared agentic frameworks.
A new source-code analysis of eleven production coding agents—including Claude Code, Gemini CLI, Aider, and OpenHands—reveals that the field has converged on custom-built runtime architectures rather than shared frameworks. The work, published on arXiv, stakes out "harness engineering" as a distinct discipline focused on the design and evolution of the runtime systems that couple large language models to tools and external systems.
What the study examined#
The researchers dissected roughly four million lines of code across Python, TypeScript, and Rust implementations from systems including Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode, and OpenClaw. They also analyzed Omnigent as a meta-harness contrast point. Each harness was mapped against seven canonical subsystems—the runtime loop, tools, context management, safety controls, orchestration, and extension surfaces—and cataloged according to minimal and maximal implementations.
Key findings#
Two striking absences stand out. Across the entire corpus, no coding harness imports a general-purpose agentic framework, and none uses vector embeddings for code retrieval; instead, the systems rely on hand-rolled async loops and deterministic approaches. The audit identified 13 cross-cutting observations and documented 29 recurring design patterns.
Skills and protocols show varied adoption: model context protocol (MCP) appears in 9 of 11 systems, while a newer protocol (ACP) ships in 6 systems and introduces a third operational role—harness hosting. The original eight systems were also re-examined after a three-month gap, revealing a longitudinal pattern in which behavioral policies are shifting from prompt-based prose toward structured configuration.
The broader shift#
The study's central claim is that coding harnesses transitioned from tool-like utilities into platform-grade infrastructure during the first half of 2026. The paper concludes with 18 design recommendations and a 90-line reference implementation to guide future harness engineering.
