- Tool descriptions are the model's actual selection mechanism — expand and differentiate them before reaching for a routing layer or tool consolidation.
- Structured MCP errors need a category, a retryability flag, and a human-readable reason — an access failure must never look like a legitimate empty result.
- Tool count and specialization both affect reliability: fewer, sharper tools per agent beat a large shared toolbox, with narrow scoped exceptions for genuine cross-role needs.
- MCP scope (project vs. user) follows who needs the server; secrets live in environment variables, never in the committed config.
Grepis for content,Globis for paths; build codebase understanding incrementally rather than reading everything upfront, and account for re-exports when tracing usage.
Key Takeaways
Estimated time: 3 minutes