Do loops actually work for anyone? I tried them for months, and even with the best specs or a well-defined goal, there is always some drift that leads to lots of wasted work that needs to be unwound. I have had much more success with single-agent sessions running overnight
Are we still talking loops or did we shift to graphs yet?
@Tim_Dettmers i have no idea what a "loop" is anymore and i'm too afraid to ask 馃槀
Some users report loop-based agent workflows succeeding for tasks like investment modeling or cron jobs, while others dismiss them as token-burning schemes benefiting providers or low-utility software.
Based on 16 visible X reactions from 51 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@Tim_Dettmers To me, loops work well when there's a clear goal and a bunch of similar experiments. You just need agents to document their findings with timestamps so they don't get stuck chasing the same details. I think with single agent it often hits contexts length issue but loops doesn't.
@Tim_Dettmers I use them for investment scenario modeling, which works because there are so many different measurable proxy variables to control stopping constraints. The list of equations for stopping constraints becomes the most important (and reusable) part of the spec.
@AllanatrixQ @eliebakouch Very interesting! I think this is a good example. Curious about any lessons you learned when you set this up
@Tim_Dettmers Of course they 'work'. How else do you get customers to burn 10x tokens? (O(n虏) btw)
Do loops actually work for anyone? I tried them for months, and even with the best specs or a well-defined goal, there is always some drift that leads to lots of wasted work that needs to be unwound. I have had much more success with single-agent sessions running overnight
@Tim_Dettmers i have no idea what a "loop" is anymore and i'm too afraid to ask 馃槀
Some users report loop-based agent workflows succeeding for tasks like investment modeling or cron jobs, while others dismiss them as token-burning schemes benefiting providers or low-utility software.
Based on 16 visible X reactions from 51 accounts; directional sample.
Ask a question below.
Published answers will appear here.
