MCP Tool Design: Engineering Practices and Trade-offs Guide

AWS ML Blog · rss · 2026-07-10

Pain Point Analysis

The article points out that poor MCP (Model Context Protocol) tool performance usually stems from tool design rather than the protocol itself, primarily due to two issues:

Bloat: Tool definitions are loaded into the LLM context on every call. Multiple MCP servers can easily exhaust the context, degrading reasoning capabilities.
Confusion: As reasoning weakens, the LLM makes poor choices (like calling the wrong tool or parameters). While rich descriptions can alleviate confusion, they may exacerbate bloat.

Optimization Strategies

The author proposes several context engineering methods to balance bloat and confusion:

Streamlined Responses & Errors: Return only essential fields by default (reducing tokens by about two-thirds) and provide clear error prompts to guide the LLM's next attempt.
Schema Constraints: Use parameter naming intuitive to LLMs, set default values, and use enums to restrict valid inputs. Keep the parameter count around 8.
On-Demand Loading: Split multi-functional tools and adopt lazy-loading discovery mechanisms (e.g., Anthropic's Tool Search Tool can reduce token consumption by 85%).
Server-Side Inference & Agentic Workflows: Offload intent parsing to an internal Agent or external LLM, transforming client LLM calls into natural language endpoints. This brings engineering performance entirely under your control.

Original post →

MCP Tool Design: Engineering Practices and Trade-offs Guide

More from coding & agent