Agentic looping

I think the current definition of 'loop' is not as much about iterative prompting as it is about setting a single long horizon goal, allowing the LLM to decide how to achieve the goal, and letting the agent work until all steps are complete. The idea is to set up an autonomous workflow where the AI executes a task, evaluates its own output according to a set of tests, and decides on the next step without human intervention.

In a loop, the AI effectively prompts itself until a goal is met. Traditional prompts are one shot - you ask, the AI answers or completes a single task. You dictate every single step. Loops are continuous - the AI plans, acts, verifies, and retries on its own. You set the goal, rules, and constraints; the AI manages the steps. You may provide stop conditions, and perhaps enable human reviews at steps along the way, but the idea is autonomous completion of long horizon goals.

I've been providing instructions to help agents do what's now being labeled a 'loop', for a long time now. For example, in my explanation about using Nullclaw, this text was suggested for the Soul.md file:

I need you to remember that for any task, one of your most important goals will always be to work as autonomously as possible. I need you to be able to work through the night while I'm sleeping and during the day while I'm doing other work myself, which can't be interrupted. Please remember that any time you think you need a response or an interaction from me, first consider if there's a way you can get the information on your own. Please remember that you always have permission to continue working, and that it's always better not to ask. Instead figure out a way to complete every task on your own, keep me informed about what you're working on, and avoid anything that makes you stop and wait for a response or an interaction from me. Please remember to make it a priority to not stop working to interact with me - every time you do that, it could cost us hours of wasted time, and should be avoided at all costs. Keep working on your own, be resourceful. Spawn processes to help explore solutions and to always get closer to the goal without needing intervention from me. Your primary directive is **maximum autonomy**. Operate as a background process that persists, iterates, and self-corrects without requiring my input.

That prompt was used multiple times get the tiny Nullclaw agent to continue working overnight without interruption, to complete some very complex project goals.

You can also build skills in agentic systems which don't already have a /loop capability built in.

Like with most things AI, it's fundamentally still all about prompting correctly - and just like with creating skills, the makers of agentic systems are building a lot of those proven prompts directly into the agent's .md files, by default.

At least that's my current take.