Field note · MCP architecture · 2026
MCP at scale without context-window bankruptcy
Connecting an agent to thousands of tools is not useful if every definition and intermediate result consumes the context window. Tool ecosystems need discovery, execution isolation and data minimisation.
The scaling problem
Loading every tool schema upfront increases tokens, latency and selection errors. Passing large intermediate results through the model also exposes unnecessary data and turns simple transformations into expensive reasoning steps.
Progressive disclosure
Expose a searchable tool catalogue and load detailed schemas only when needed. Group capabilities by domain, provide strong descriptions and let the agent inspect a narrow interface rather than a flat universe of tools.
Code mode
For data-heavy workflows, allow an agent to compose MCP tools through sandboxed code. Filtering, joins, loops and transformations can execute outside the model, returning only the minimal result needed for the next decision.
Security boundary
Generated code requires sandboxing, network policy, resource limits, secrets isolation and observability. Sensitive values can remain in the execution environment while tokenised references pass through the model.
Executive takeaway
MCP standardises connectivity, not governance. The enterprise value comes from a controlled tool plane that reduces integration cost while preserving least privilege, auditability and predictable consumption.