Korva is a scroll of the internet’s best ideas.
One shot is never enough
Claude Code is a harness, and it works the way a harness inside any product would. A prompt goes in and an answer is generated. The model thinks a little more, calls itself to get more context, and may spawn a sub-agent. It will likely ask to run tools, such as reading files or writing a unit test to see whether a bug exists, and then runs that test and feeds the result back into its context.
More capable models tend to run more tools, not fewer. A PhD-level expert asks more questions than a less capable one, since a right answer depends on specifics. ChatGPT in December 2022 gave a single, somewhat wrong answer. Now a model may ask three clarifying questions, which catches ambiguity before it produces a misleading reply.
The only way to skip loops is a one-shot prompt that returns a golden answer, which would be AGI. Without that, the model has to be steered with thoughts, tools, and context, and that is always some form of loop.
Why do capable models ask more questions?
A request with missing details has many valid answers, and a model that notices the gap can ask before committing to one. In agent work a wrong assumption can spread through many later edits, so a question early on costs far less than a repair afterward. Information retrieval researchers have studied when a system should ask a clarifying question instead of answering for years.
Why do tool results improve a model's answers?
A language model generates text from patterns and cannot verify its own claims. Running a unit test returns a result from the actual code, which enters the context as evidence and replaces a guess with an observation. A failing test or a compiler error is unambiguous feedback the model can act on. SWE-bench, released in 2023, scores models on resolving real GitHub issues, with success decided by running the project's tests.
How have chat assistants changed since 2022?
ChatGPT opened to the public in late 2022 as a text responder with no tools. OpenAI added function calling to its API in June 2023, which let models request outside actions in a structured form. In September 2024, OpenAI's o1 model spent additional tokens reasoning before it answered, which added another layer of looping inside a single reply.
More from The Peel with Turner Novak
