A production MCP pattern hides multi-agent orchestration behind a single thick tool
anilkr84 · reddit · 2026-07-26
Moving orchestration into the MCP server hides a multi-agent system behind one tool
The author describes a production pattern for MCP-based systems: instead of making the client LLM orchestrate a long chain of tools, move the reasoning and routing into the server and expose only one thick tool such as askagent(goal).
The three MCP generations
- Gen 1: thin stateless tools; the client handles all planning and state.
- Gen 2: tool plus skillpack; instructions improve usage, but orchestration still lives in the client.
- Gen 3: server-side orchestration; one call in, one answer out, with intermediate state kept on the server.
What this fixes
- Token cost stays flatter as tasks get deeper.
- Less goal drift, because the client only holds the question and answer.
- Smaller schema overhead.
- Routing can be handled by a domain-specific brain that adapts at runtime.
- The same tool can work across Claude, ChatGPT, or Codex.
Tradeoffs
- Higher latency per call.
- Less transparency and mid-chain steering.
- You are effectively running an agent system in production, so evals and observability become mandatory.
The author says their practical setup is a hybrid: one thick reasoning tool plus a few thin control tools for simple list/select operations.
Related event: Architecture Insight: Hiding Multi-Agent Orchestration in a Single MCP Tool(2 posts)→
More from coding & agent
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26
- Hugging Face ships hf CLI updates: deduped cache saves 19GB, faster buckets, retryable GPU jobs — yb2698 · 2026-10-26