Loops are all the rage these days, but there are some primitives that I think needs to be understood first before going to the next level of abstraction which agentic loops unlocks. What is an agent, what is context window and its limit, what makes it possible to unlock long running task, the minimal setup needed.
The post basically talks about a dumb (minimal) setup and its dumb because its just a loop over claude code, but its simple and it works it what is makes is so interesting and when done right unlocks another level of productivity where one is not limited to their token budget.
I think loops serve a valuable purpose when the work needs to be repeatable and there’s a certain measure of precision involved. I do think, for more abstract problems, free thinking and open question prompting can really explore possibilities not otherwise thought of.
Context rot is always going to be a challenge and it evolves over time. It’s interesting to see the dumb harness idea, i do believe in taking large concepts and workshopping them at a smaller level to prototype and iterate quickly.
"I do think, for more abstract problems, free thinking and open question prompting can really explore possibilities not otherwise thought of." But the question, is can this itself be engineered in a loop. For example, I have a specific way of enquiry when I do one one one with claude code for example for such brain storming, but if that line of thoughts can be loop engineered based on my standards and taste with some level of visibility (an interesting design problem to solve) so I can have confidence is that system that can explore new ideas and I am still involved but not too much hand holding that session and the loop can take that forward. I am struggling to put into works but this needs to be more dissected and understood. Loops engineering needs to be more understand and what kind of possibilities they unlock.
That's why you need a verifier and an acceptance criteria, at the end of each iteration and going one more level using a different model for that QA purpose different from the model that does the specification / implementation.
Hi HN,
Loops are all the rage these days, but there are some primitives that I think needs to be understood first before going to the next level of abstraction which agentic loops unlocks. What is an agent, what is context window and its limit, what makes it possible to unlock long running task, the minimal setup needed.
The post basically talks about a dumb (minimal) setup and its dumb because its just a loop over claude code, but its simple and it works it what is makes is so interesting and when done right unlocks another level of productivity where one is not limited to their token budget.
Have a read and let me know your thoughts.
*not limited -> only limited
I think loops serve a valuable purpose when the work needs to be repeatable and there’s a certain measure of precision involved. I do think, for more abstract problems, free thinking and open question prompting can really explore possibilities not otherwise thought of.
Context rot is always going to be a challenge and it evolves over time. It’s interesting to see the dumb harness idea, i do believe in taking large concepts and workshopping them at a smaller level to prototype and iterate quickly.
"I do think, for more abstract problems, free thinking and open question prompting can really explore possibilities not otherwise thought of." But the question, is can this itself be engineered in a loop. For example, I have a specific way of enquiry when I do one one one with claude code for example for such brain storming, but if that line of thoughts can be loop engineered based on my standards and taste with some level of visibility (an interesting design problem to solve) so I can have confidence is that system that can explore new ideas and I am still involved but not too much hand holding that session and the loop can take that forward. I am struggling to put into works but this needs to be more dissected and understood. Loops engineering needs to be more understand and what kind of possibilities they unlock.
Loops are interesting no doubts. However the bottleneck is still the QA
That's why you need a verifier and an acceptance criteria, at the end of each iteration and going one more level using a different model for that QA purpose different from the model that does the specification / implementation.