How Much Should You Tell Your AI Agent? (www.raymondyxu.com)

Anthropic’s recent Claude Sonnet 4.5 launch — billed as capable of working autonomously for “30 hours” — sparked Raymond Xu’s practical guide on how to prompt long-running LLM agents. Xu argues that designers must balance explicit rules and high-level goals: too many enumerated steps make modern models rigid and brittle, while too-vague principle-only prompts leave agents unable to handle edge cases. He uses a detective analogy to show why agents need freedom to adapt, and critiques Anthropic’s own example prompts (too specific, just-right, too vague) to highlight what to keep: explicit edge-case rules where intuition fails, plus concise goal statements and company principles in the system prompt. Technically, Xu points out error compounding in multi-step tool use — e.g., a 10% mistake rate per tool call yields only ~35% success after ten rigid steps (0.9^10 ≈ 0.35) — so long-horizon agents should be goal-aligned and able to detect and course-correct rather than strictly follow a checklist. Practical takeaways: put a short, clear mission and values in the system prompt, include a few concrete examples or escalation rules that can’t be inferred, and favor a response framework over exhaustive step lists. The result: agents that tolerate occasional wrong steps but steadily progress toward objectives, enabling more robust, longer-running automation.