Designing MCP for the Age of AI Agents

Why It Matters

By dramatically reducing context overhead, Harness v2 lets LLM agents reason more effectively, accelerating AI‑driven DevOps workflows while maintaining governance and safety. This sets a new efficiency benchmark for enterprise‑scale MCP implementations.

Key Takeaways

• v2 reduces tools from 130+ to 11
• Context usage drops to ~1.6% of 200K tokens
• Registry dispatch supports 125+ resource types without new tools
• Safety features include confirmations and read‑only mode
• Enables multi‑MCP stacks under IDE tool caps

Pulse Analysis

The rapid adoption of AI agents in software delivery has exposed a hidden bottleneck: tool sprawl. Traditional MCP servers expose a separate tool for every API endpoint, forcing large language models to allocate a sizable fraction of their context window to tool definitions. When that overhead approaches a quarter of a 200 K‑token window, the model’s attention dilutes, leading to slower reasoning and higher error rates. Researchers and practitioners alike have highlighted that trimming irrelevant context restores the model’s focus on the core problem, a principle that underpins Harness’s redesign.

Harness’s MCP v2 replaces the endpoint‑per‑tool paradigm with a registry‑based dispatch model. A single set of eleven generic tools—describe, list, get, create, update, delete, execute, search, diagnose, status, and ask—acts as a stable vocabulary. Behind the scenes, a declarative registry maps each resource type to its underlying API calls, handling scope, pagination, and response extraction automatically. This architecture scales effortlessly: adding new services merely involves inserting a new registry entry, not expanding the tool surface. The approach also embeds safety mechanisms, such as human‑in‑the‑loop confirmations and read‑only modes, ensuring that powerful automation remains governed.

The broader impact reaches beyond Harness. By demonstrating that a rich platform can be exposed to LLM agents with minimal tool definitions, the v2 server establishes a template for other enterprises grappling with extensive APIs. Developers can now combine multiple MCP servers within IDE limits, preserving up to 98 % of the context window for actual reasoning. This efficiency translates to faster debugging, streamlined onboarding, and more reliable AI‑assisted operations, positioning AI‑augmented DevOps as a practical reality rather than a theoretical concept.