Anthropic on Why Non-Coding Agents Fail

Anthropic analyzed millions of real human-agent tool calls and found that about 50% of agent activity on the platform is concentrated in software engineering. In contrast, non-coding fields such as sales, marketing, finance, and law remain in single-digit shares.

The core reason agents fail in non-coding environments is not a lack of the model's reasoning ability, but rather the unstructured nature of data.

• Strengths of coding agents: They operate in structured environments like Git repositories, benefit from strict syntax rules, work within deterministic terminal environments, and receive clear compiler messages when errors occur.
• Weaknesses of non-coding agents: Unstructured web data—such as competitor price fluctuations, improperly formatted data, and countless HTML tags and tracking scripts that fill up the context window—degrades model performance and causes hallucinations.

Ultimately, building successful non-coding agents depends less on the model itself and more on infrastructure work such as data preprocessing, reliable scraping, and clean input management.