Ship more software
with smaller teams.

Give an agent a large, loosely defined goal and let it run, and it drifts. Long-running tasks lose the thread. Genuinely complex problems do not survive being held in one long session. And because nothing is checked along the way, you find out at the end — when the work is finished, confident, and wrong in a way nobody has looked for.

The usual response is to run it again. A better prompt, a longer leash, another few hours. But iteration is not a method, and repeating an unplanned process mostly buys a more expensive way to arrive at the same place. Endless loops do not fix anything, because the missing thing was never more attempts. It was a plan.

What works is deliberate and curated. Decide the architecture before anything is built. Break the requirement into pieces small enough to check on their own. Write down what “done” means before the work starts, so passing is a fact rather than an opinion. Verify each piece as it lands, and send failures back instead of letting them accumulate into the next task. Keep a person on the two decisions that carry the most consequence.

That structure is the whole of what LFG is. It replaces the loop with a plan, and the plan is what makes the output something you can hand to a client.