Tony argues smarter AI models will still need more loops, not fewer
The alternative to loops is a perfect answer in one try, which is AGI, says Inngest co-founder Tony.
The Peel with Turner Novak · The Hidden Layer Every AI Agent Runs OnTony, the co-founder of Inngest, a company that sells the plumbing for running software in steps, was answering a skeptical question. If AI models keep getting better and cheaper, why bother with all the custom orchestration around them, the loops and scaffolding that feed them context and check their work.
His answer was blunt. "One shot is never enough," he said.
By one shot, Tony means the fantasy where you ask an AI one question and get back a perfect answer with no follow up. The only alternative to loops is what Tony called "this golden answer, which is literally AGI," a hypothetical AI that can do any intellectual task a human can. Until then, he believes, someone has to build the harness, the code that lets a model think, call tools, and try again.
His example is Claude Code, the coding assistant. Tony describes it as a harness itself. It takes a prompt, thinks, calls itself to get more context, sometimes starts a smaller helper agent, and asks to run tools. It might read files, write a small test to check if a bug exists, run that test, and then feed the results back in to decide what to do next.
That does not go away with smarter models, in his view. It increases. A more capable model asks more questions, not fewer. Even a PhD level human expert, faced with a vague request, will ask for clarification, and ask more than a weaker model would.
He has watched that shift in AI. In December 2022, ChatGPT would just give an answer, and it was often kind of wrong. Today, the model will ask three qualifying questions first. The user realizes the request was ambiguous, supplies the missing detail, and avoids getting a confident and misleading answer. For Tony, that extra thinking and clarifying is the point, and it is why the loops have to stay.
Questions people ask
What exactly is a harness or agent loop?
A harness is the surrounding software that manages an AI model while it works. It sends the prompt, gives it tools like web search or file access, keeps track of what has happened, and loops back for another try until the job is done. A loop is that back and forth cycle of thinking, acting, seeing the result, and trying again.
What does AGI mean here?
AGI stands for artificial general intelligence, a commonly used term for a future AI that could handle any intellectual task a skilled human can, without special setup. Tony uses it to mean a system smart enough to give the right answer from a single prompt, with no extra context or retries. No such system exists today.
If smarter models run more tools and ask more questions, do they cost more to use?
They often do per task, because each extra step uses more computing and more fees for input and output. Developers accept that when the extra steps produce a more reliable answer. Much of current AI engineering is trying to keep quality high while shifting some of those steps to smaller, cheaper models.