Small Loop vs Big Loop

🔁 Small Loop (Inner Loop)

Every agent has a tiny cycle running inside it:

send context → model asks for tools → run tools → add results → repeat

This stops when the model itself decides it's done. Problem: nothing checks whether the model is actually right — it's just judging its own work. That's why agents confidently say "Done! All fixed" without ever running the tests.

Fix = Outside stops — conditions that don't depend on the model's own opinion:

• A checked condition (prove the work with a real test)
• A limit (max number of tries)
• A no-progress check (stop if nothing's improving)
• A separate checker (a second process grades the work)

🧭 Big Loop (Outer Loop)

The small loop is one worker doing one task. The big loop is the manager — it decides:

• Which task to give
• When to start
• How to grade the result
• What to remember for tomorrow

One full run of the small loop = one "beat" of the big loop.

🏗️ The 4 Layers (each nested inside the next)

| # | Layer | What it covers | |---|-------|-----------------| | 1 | Prompt engineering | The words you send | | 2 | Context engineering | Everything the model sees in one turn | | 3 | Harness engineering | The code around the model (running tools, handling errors) — where the small loop lives | | 4 | Loop engineering | The outer cycle: what the system works on, when it starts, how it knows it's done |

Each layer stops a different kind of failure:

• No context → the model guesses
• No harness → you are the only checker
• No loop → the schedule is still on you

🎯 Key Takeaway

The useful question isn't "is my prompt good enough?"

It's: "Which of these layers am I still doing by hand?"

The small loop (the inner loop). Inside every agent, there is a tiny cycle of code. It works like this: send the context to the model → the model asks to use tools → run the tools → add the results to the context → repeat. When the model stops asking for tools, the cycle ends.

The small loop stops when the model itself decides it is finished.