Most Agent Failures Are Context Failures — Rostislav Melkumyan, Sanity

Effective AI agent design for content operations relies on managing "context shape" rather than simply increasing the number of available tools. By prioritizing an "overview first, detail on demand" approach, agents can navigate complex schemas and large datasets without exceeding context limits. Utilizing "sets"—reusable handles to query results—allows agents to reason over thousands of documents efficiently. Reliability is further enhanced by allowing models to draft queries while using code to validate and repair them, ensuring structural integrity. This strategy avoids the pitfalls of multi-agent systems, which often suffer from context loss, by maintaining a single, strong agent supported by specialized tools. Ultimately, crisp, behavior-driven instructions and clear tool boundaries outperform complex, fragmented agent architectures, enabling scalable and trustworthy content management within systems like Sanity.