Code Execution with MCP: Fix Tool Token Bloat (Adam Jones, Anthropic)

Описание: As MCP (Model Context Protocol) adoption grows, many teams run into the same problem: too many tools, too much context, and too many tokens burned before real work even starts.

In this talk, Adam Jones walks through why traditional MCP clients struggle with:

Tool definition bloat from large schemas and long descriptions
Tool result bloat when large outputs are copied back into model context

He then introduces a code execution approach for MCP that treats MCP servers like SDKs. By generating typed stubs and letting models write and execute code in a sandboxed environment, MCP clients can:

Load only the tools they actually need (progressive disclosure)
Pass large intermediate results through variables instead of context
Use real control flow (loops, conditionals, batching, polling)
Reduce cost, latency, and context usage at scale
Build reusable “skills” and composable workflows over time

The session includes practical examples—from copying large documents between systems to polling services efficiently—showing how code execution enables more capable, reliable, and scalable MCP-based agents.

If you’re building MCP clients or servers, this talk offers a clear look at where MCP is headed and how code execution unlocks the next level of agent behavior.