MCP Changed How I Think About AI Agent Orchestration
MCP Changed How I Think About AI Agent Orchestration
I spent weeks evaluating agent orchestration frameworks - LangGraph, CrewAI, custom state machines. Then I tried passing a simple JSON object between steps. It worked better than all of them.
The Over-Engineering Trap
Most orchestration frameworks solve problems you do not have yet. They add abstractions for parallel execution, conditional branching, error recovery trees, and retry policies. You write more framework configuration than actual agent logic.
The real need is simpler: pass context from one step to the next, track what has been done, and know when to stop.
JSON State - The Simple Version
A single JSON object holds everything the agent needs to know:
{
"task": "Send weekly report email",
"steps_completed": ["gathered_data", "formatted_report"],
"current_step": "composing_email",
"context": {
"report_data": "...",
"recipient": "team@company.com"
},
"errors": []
}
Each step reads the state, does its work, updates the state, and passes it forward. No framework needed. The state object is the entire orchestration layer.
Why This Works with MCP
MCP (Model Context Protocol) gives agents access to tools. Each tool call is a step. The agent reads the state, decides which tool to call, updates the state with the result, and moves on. The LLM itself handles the orchestration logic - deciding what comes next based on the current state.
You do not need a framework to tell the LLM what to do next. You need a clean state object so the LLM knows what has already happened.
When You Need More
If you are running dozens of agents in parallel with complex dependencies, a framework makes sense. If you need durable execution that survives process crashes, you need proper workflow infrastructure.
But for most single-agent workflows - which is most workflows - a JSON state object and a list of available tools is the entire architecture. Start there. Add complexity only when the simple approach fails.
Fazm is an open source macOS AI agent. Open source on GitHub.