Building Agents for Production Systems with MCP
Anthropic's guide to connecting AI agents to production systems, making the case for MCP as the standard integration layer. The key framing: "Agents are only as useful as the systems they can reach." Three approaches compared (direct API calls, CLIs, MCP), with MCP winning on scalability -- one remote server reaches every compatible client. MCP SDK downloads hit 300M/month, up from 100M at the start of the year.
Key Quotes
"Agents are only as useful as the systems they can reach."
"Group tools around intent, not endpoints -- fewer, well-described tools consistently outperform exhaustive API mirrors."
"As production agents move to the cloud, MCP becomes the critical layer, and it's the one that compounds."
Key Themes
The "group tools around intent" pattern is the most actionable design advice. A single create_issue_from_thread tool outperforms four separate tools for the individual steps. Cloudflare's approach -- 2,500 endpoints through just 2 tools that accept code -- is the extreme version. This is API design for AI consumers, not human consumers, and the principles are different.
The context efficiency section (Tool Search reduces tool-definition tokens by 85%+, Programmatic Tool Calling reduces tokens by ~37%) provides concrete optimization targets. These aren't marginal improvements -- they're the difference between agents that run out of context and agents that don't.
Critical Analysis
This is Anthropic's official position paper on MCP, so it's both authoritative and self-interested. They built MCP, they're pitching MCP. The arguments are sound but the alternatives (direct APIs, CLIs) are underweighted. For many use cases, a CLI wrapper is simpler, faster to build, and good enough. The MCP advantage is strongest for cloud-hosted agents that can't access a local filesystem -- which is Anthropic's product direction (Claude Cowork, Managed Agents).
The Skills + MCP pairing section is forward-looking: MCP provides tool access, skills provide procedural knowledge for using those tools. The upcoming protocol extension to deliver skills from MCP servers would create a single distribution mechanism for both capabilities and playbooks -- a meaningful step toward agent composability.
Willison's Stateless MCP provides the independent-practitioner complement: where this guide argues from Anthropic's platform perspective, Willison validates the stateless spec by actually building against it — shipping three tools (mcp-explorer, datasette-mcp, llm-mcp-client) in a week, with datasette-mcp succeeding on its fourth attempt only because the stateless redesign removed the complexity that had killed three prior versions.
For practitioners building agent systems, this is essential reading alongside Awesome Agentic Patterns (the pattern catalogue) and Two Kinds of User Are Emerging (the adoption context). Sierra's The MCP Gateway Iceberg provides the most detailed production case study to date — validating the CLI-over-MCP pattern (Lesson 6), documenting the "80% of a workflow = 0% of the value" coverage problem, and describing a multi-pass cross-customer data guard that operationalizes the tool-design principles laid out here.
AMD's Execution-Free Agentic Program Repair provides the strongest empirical validation of this article's tool-design principles: their ablation study shows that exposing the full GitHub/Jira MCP toolsets (vs. a pruned set of 6 essential tools) drops localization accuracy from 70.5% to 53.6%, and eliminating response filtering (HTML cleanup, metadata stripping) causes a catastrophic collapse to 23.2% localization with 46.4% no-patch rate. Tool curation and response normalization aren't optimizations — they're the difference between a working system and a broken one.
Sources: Building Agents for Production Systems with MCP Last updated: 2026-05-14