-1
I do this everyday for technical writing (and it absolutely works! (tldr.nettime.org)
submitted 3 weeks ago by to c/programming@fedia.io
3 comments fedilink hide all child comments
 

I do this everyday for technical writing (and it absolutely works!)"Adversarial Code Review is a verification pattern where a distinct AI session—the Critic Agent—reviews code produced by the Builder Agent against the Spec before human review.This extends the Critic (Hostile Agent) pattern from the design phase into the implementation phase, creating a verification checkpoint that breaks the “echo chamber” where a model validates its own output.The Builder Agent (optimized for speed and syntax) generates code. The Critic Agent (optimized for reasoning and logic) attempts to reject it based on spec violations."asdlc.io/patterns/adversarial-#AI #GenerativeAI #LLMs #Programming #SoftwareDevelopment #CodeReview

you are viewing a single comment's thread
view the rest of the comments
[–] 1 point 3 weeks ago (1 child)

@remixtures i don’t call it that but work that way anyway - i design and discuss with chatbot higher model - abd implement with builder in terminal - but have each one teview the other - but the builder gets most decisions on implementation as it sees the code base directly _

  • permalink
  • fedilink
  • source
  • hideshow 1 child comment
    •
  • [–] [S] 1 point 3 weeks ago

    @zerowan Yes, I use Claude Opus 5.5 via Claude Code to check the source code of the repositories and then write the documentation. But then I use GPT 6.1 Sol to review the text produced by Opus. This is great because you can never guess which model makes the best findings. In any case, lately Claude seems to accept the majority of the suggestions and recommendations provided by GPT. I don't even know how can someone say that the outcome of such an itirerative process is just slop. Yes, it takes a lot of time and persistence. Yet, for the great majority of scenarios, the quality of the final output is always better than it would be without using AI agents/LLMs. This will, of course, not work as expected if there's no human being in the loop to make use of his judgement and critical thinking skills when agents seem to lack the proper context.

  • permalink
  • fedilink
  • source
  • parent