An academic study found developer-written instruction files for AI agents reduce agent-introduced bugs by 35-55%. In contrast, instructions generated by an LLM actually decrease task success rates and increase inference costs by over 20%. This highlights the critical value of human judgment in steering AI systems effectively.
A prompted instruction like "never do X" is merely a probabilistic suggestion to an AI model and can fail. For critical rules, use 'hooks'—deterministic code that fires on specific events. This provides a guarantee of enforcement for actions that must always or never happen, a reliability that prose-based prompts cannot match.
The key function of subagents in multi-agent systems is to isolate context. By running side tasks like log analysis in a fresh, separate context window, they prevent intermediate results from cluttering the main agent's session. Only the final summary is returned, reducing token costs and improving the main agent's focus.
Moving beyond proprietary files like CLAUDE.md, the AGENTS.md convention is emerging as an open standard for instructing AI agents. Stewarded by the Linux Foundation and backed by OpenAI, Google, and Microsoft, it allows teams to create a single source of truth for project instructions that works across multiple coding agent platforms.