Anthropic analysis explains why non-coding AI agents fail

View on Reddit

Anthropic analyzed millions of real human-agent tool calls across their public API. They found that software engineering makes up roughly 50% of all agentic activity.