Getting Started with Agentic AI

Sample post — swap for your own. Written from the perspective of a frontend engineer moving into AI engineering.

Coming from a frontend background, the mental model shift for agentic AI is smaller than it looks. An agent loop is just a state machine where one of your "reducers" happens to be a language model.

The core loop

At its simplest, an agent is:

while (!done) {
  const response = await model.respond(context, tools);
  if (response.toolCalls) {
    context.push(await runTools(response.toolCalls));
  } else {
    done = true;
  }
}

Everything else — RAG, MCP, planning — is about what you put in context and which tools you expose.

Where MCP fits

The Model Context Protocol standardizes how you expose tools and data to a model. Instead of hand-rolling function schemas, you connect to MCP servers that advertise their capabilities.

What surprised me

• Grounding matters more than cleverness. Good retrieval beats a clever prompt.
• Determinism is a feature. Structure the loop so you can replay and debug it.
• Evals are the real work. Shipping is easy; knowing it's good is hard.

Next

I'm building a small developer assistant on this stack. More notes to come.