Lesson 5 of 7

The agent loop & the 2026 agent stack

8 min read

Strip away the frameworks and an agent is one small loop repeated. What's actually in it — and what's the rest of the stack around it?

The loop: think → act → observe

Every agent runs the same cycle. Think: the model decides the next action. Act: it calls a tool. Observe: the result comes back into context. Then it loops — think again on what it just learned — until the goal is met or a limit stops it. That's it. The "magic" is the same three steps repeating on fresh observations.

An agent is a loop: think → act → observe, repeating on each new result until done. Reacting to the observation is the whole difference from a one-shot answer.

The stack around the loop

In production, the loop is the small part. Around it sits the stack: a model that plans, a tool layer (often MCP) it acts through, memory for state beyond the window, orchestration to run and retry steps, and observability — tracing, evals, cost and token limits — so you can see what it did and stop it when it drifts. Frameworks assemble these; the loop stays the same.

The 2026 agent stack is the loop plus model, tools, memory, orchestration, and observability. Frameworks come and go; those layers are the durable shape.

A framework is optional. You can write the loop as a plain while over a chat call — think, run the tool, append the result, repeat. Reach for a framework when orchestration, retries, and tracing become the hard part, not before.