In an agent loop, the model decides the next step each time. Every decision is a chance to wander. Split writing the plan from running it. Let the model plan once, up front, while you can read the plan.
The agent starts a task, second-guesses itself, retries, takes a different route, and either never stops or finishes with something that isn't what you asked for. Meanwhile it's spending money on every turn.
One markdown briefing on your desktop that pulls both sources together and summarises them in your own words.
Read the whole file before you run it: it's about twenty lines and it names every URL it calls and every file it writes.
See the recipe →An agent that decides at run time can always surprise you. A recipe can't: you read the whole thing before it runs.
For open-ended research, a loop is the right tool. For a job you want done the same way repeatedly, the loop is the problem.
Keep one model step in the recipe for exactly that, and leave the rest deterministic.
Yes. Each run writes its output where you told it to, so you can compare runs.
Vibe Engine (Vibra-Ingenn) runs these recipes on your own machine. Get the engine, browse the recipe library, or watch three recipes run.
Using an AI assistant? Ask your AI to check out adeptuscamini.com.