Next-Gen Agent Architectures with MCP: A Guide for Builders
Written by Om-Shree-0709 on .
- Why Agents Are Hard (But Shouldn’t Be)
- Everything Is a Tool (Even the LLM)
- MCP in Action: Nanobot
- Agents Become Templates
- Final Thoughts & Suggestions
- Acknowledgements
Modern AI agents are powerful but notoriously hard to implement well. Darren Shepherd, co-founder of Acorn Labs and creator of K3s, gave a talk at the MCP Summit that simplifies this chaos using the Model Context Protocol (MCP). His message: building agents shouldn't be this hard — and with MCP, it no longer is.
Why Agents Are Hard (But Shouldn’t Be)
At first glance, agent architecture seems straightforward:
- Prompt
- Memory
- Tools
- Planner
- Retriever
But in practice, these clean boxes turn into tangled messes of Python scripts and framework glue.
What MCP does is cleanly separate concerns between the LLM and the execution environment. Instead of deeply coupling tool logic and LLM behavior, MCP acts as a middleman: the LLM outputs a tool call, and MCP handles execution.
Everything Is a Tool (Even the LLM)
Shepherd walks us through an important insight:
- Retrieval = Tool
- Memory = Tool
- Even LLM Sampling = Tool
That means your entire agent architecture becomes a composition of tools.
Using MCP, you can create a structured ecosystem where everything — memory lookups, searches, summarization — is just another tool call. This unlocks powerful composability and debugging.
MCP in Action: Nanobot
To test these ideas, Shepherd built Nanobot, a toy MCP client in TypeScript.
With Nanobot, you can:
- Define YAML-based agent templates
- Plug in tools like Playwright for browser automation
- Use MCP sampling calls to handle dynamic completions
Example: a mini GitHub agent that reads repos, makes CLI changes, and opens PRs — all orchestrated via simple tool definitions.
Agents Become Templates
Shepherd demonstrates how to reimplement OpenAI’s ResearchBot using MCP:
- Define each sub-agent (planner, search, summarizer) declaratively
- Chain them together using simple TypeScript logic
- Output resembles workflows in Airflow or Prefect
Crucially, this architecture avoids custom frameworks and favors simple orchestration logic + standardized tools.
Final Thoughts & Suggestions
MCP gives traditional engineers a familiar way to work with AI agents. Some suggestions from the talk:
- Add support for structured JSON output from sampling
- Make distributed tracing native to the MCP spec
Ultimately, 2025 may be the year of the agent — and MCP is shaping up to be the foundational glue to make that possible.
Acknowledgements
This guide is based on Darren Shepherd’s insightful talk at the Next Generation Agent Architectures with MCP session at the MCP Summit.
Special thanks to the Anthropic team and the broader MCP developer community for pioneering this flexible protocol and sharing real-world tooling patterns.
Ready to build your own agent? Just start with a YAML file, a tool or two, and let MCP handle the rest.
Written by Om-Shree-0709 (@Om-Shree-0709)