I've had really great results with cheap flash agents using a system that has a parent agent that spawns 2 sub agents. They remain active throughout the whole session and don't lose context.
The 2 sub agents get the exact same prompt and tools as each other and they independently answer but before responding back up to the parent they exchange their findings which must be backed up by a citation and then debate their findings and citations repeatedly until they can agree on correctness of all points. They are also heavily encouraged to never assume and always verify with the prompt and tools and guards where they can be realistically enforced.
It churns through tokens but tokens are crazy cheap with some of the plans out there. Could Opus one shot most of those same queries in like 1/4 time? Yep, but that's not the point for me.
DeepSeek Flash for parent then DeepSeek Flash and GLM 5.3 flash as the debating subagents gives best overall results in my experience.
You could replace the exchange at the end with an auditor subagent that takes the two results and merges them where they are different. If you can run tests, this is N Versions or cleanroom software development. Cheap tokens make it very viable where it wasn’t before with expensive humans.
I played with that idea too but found I got better results with them debating each other. Not sure why but I assume the session context for each of them helps a bit.
For context I do equal parts sysadmin/devops and having them write software
I find judges are better, there is some theory behind independent error occurrences that allows you to get better results. It really depends on how the problem is structured though.
I don't understand how this submission title ("Claude was faster and cheaper. We took the scenic route anyway") has anything to do with the article. I read the entire article and I'm still confused.
"Verify, Then Vote", "When five models disagree", or both ("When five models disagree: Verify, Then Vote") would probably have made a better one.
I've had really great results with cheap flash agents using a system that has a parent agent that spawns 2 sub agents. They remain active throughout the whole session and don't lose context.
The 2 sub agents get the exact same prompt and tools as each other and they independently answer but before responding back up to the parent they exchange their findings which must be backed up by a citation and then debate their findings and citations repeatedly until they can agree on correctness of all points. They are also heavily encouraged to never assume and always verify with the prompt and tools and guards where they can be realistically enforced.
It churns through tokens but tokens are crazy cheap with some of the plans out there. Could Opus one shot most of those same queries in like 1/4 time? Yep, but that's not the point for me.
DeepSeek Flash for parent then DeepSeek Flash and GLM 5.3 flash as the debating subagents gives best overall results in my experience.
You could replace the exchange at the end with an auditor subagent that takes the two results and merges them where they are different. If you can run tests, this is N Versions or cleanroom software development. Cheap tokens make it very viable where it wasn’t before with expensive humans.
I played with that idea too but found I got better results with them debating each other. Not sure why but I assume the session context for each of them helps a bit.
For context I do equal parts sysadmin/devops and having them write software
I find judges are better, there is some theory behind independent error occurrences that allows you to get better results. It really depends on how the problem is structured though.
[flagged]
I don't understand how this submission title ("Claude was faster and cheaper. We took the scenic route anyway") has anything to do with the article. I read the entire article and I'm still confused.
"Verify, Then Vote", "When five models disagree", or both ("When five models disagree: Verify, Then Vote") would probably have made a better one.