← All episodes

2026-08-02 · Killing Coding Agent Slop With Adversarial Self-Play cover art

2026-08-02 · Killing Coding Agent Slop With Adversarial Self-Play

Show notes

BRINE — 2026-08-02 · show notes

Guest: the tooling optimist (a fictional archetype).

Claims are paraphrased and attributed; nothing is read verbatim. Where a thread disagreed with the article, the show surfaces the disagreement.

Segments

  1. Killing Coding Agent Slop With Adversarial Self-Play
  • Source: https://usetelos.ai/blog/killing-coding-agent-slop
  • Discussion: https://lobste.rs/s/k5ibd3
  • Topic: AI Engineering · interest 75
  • The article explores using adversarial self-play—specifically leveraging weaker models as verifiers for stronger models—to filter low-quality AI code generation. The discussion adds nuance by debating whether this approach is superior to human-authored code with AI review, highlighting the inherent trade-offs in search and verification cycles.
  1. Your AI Is Not a Tool
  • Source: https://theconvivialsociety.substack.com/p/your-ai-is-not-a-tool
  • Discussion: https://lobste.rs/s/6vsam1
  • Topic: AI Ethics & Philosophy · interest 75
  • The article argues that AI is not a neutral tool but an 'environment' that subtly reshapes human cognition and judgment, even for diligent users. The discussion thread adds value by framing the author's argument within historical Marxist concepts of base and superstructure, while also debating the psychological distinction between traditional tools and addictive digital technologies.
  1. Why Developers Use LLMs to Write Blog Posts
  • Source: https://writethatblog.substack.com/p/report-llms-tech-blogs
  • Discussion: https://lobste.rs/s/gvydfo
  • Topic: developer blogging · interest 75
  • A survey-based analysis of 181 developers revealing how LLMs are used for blog content. The findings suggest that while inexperienced writers use LLMs to overcome intimidation, experienced writers use them primarily as a starting point, with a significant majority performing substantial editing or rewriting.

Transcript

Transcript. Paraphrased; sources in notes.md.

HostWelcome to the show. It is August 2nd, 2026, and we have a packed slate today. We are looking at adversarial self-play in coding agents, the philosophical debate over whether AI is a tool or an environment, and finally, a deep dive into why developers are using LLMs to ghostwrite their blogs. Samantha, I know you have been spending a lot of time in the weeds with agentic workflows lately. You ready to dive in?

GuestI am so ready, Daniel. And honestly, looking at that first topic about adversarial self-play, it hits home. I spent half of last Tuesday trying to debug an agent loop that was just hallucinating library imports like they were going out of style. The idea of using a secondary, weaker model as a gatekeeper for code quality? It is exactly the kind of pragmatic hack I love.

HostLet us start there then. Over on Lobsters, the community is digging into a piece about killing off coding agent slop by having models verify each other. The author claims that because verification is computationally easier than generation, you can use these weaker verifiers to filter out the junk. A Lobsters user called bitshift points out that this is an old concept from mathematical prover games, but they also bring up a funny point. They ask, if we are admitting humans are still better at writing code than the strongest models, why aren't we just using the AI to clean up human code instead of the other way around?

GuestSee, bitshift is onto something, but they are forgetting the friction of human output. Look, I want my agents to be my force multiplier, not just a proofreader for my own typos. The beauty of the agent-on-agent approach is that it moves at machine speed. I can spin up a sandbox environment, run the tests, and if the verifier model flags it, the agent tries again before I even see the pull request. It is not about reaching perfection, it is about keeping the sludge out of my codebase. A user called cajually noted that we are still in the dark on exactly how much sloppy code a project can sustain before it starts to rot. That is the real metric I am watching.

HostThat brings us to our second topic, which takes a bit of a turn into the philosophy of it all. There is a piece arguing that we need to stop calling AI a tool, and start seeing it as an environment that subtly shapes our cognition. The author is frustrated by the "it is just a tool" framing. Over on Lobsters, a user named Yogthos argues that this is just rehashing Marx’s concepts of base and superstructure. They say the author is acting like they found a grand revelation, when really, industrial machinery has been shaping our social consciousness for centuries.

GuestI think the author is trying to be profound, but Yogthos hits the nail on the head. Calling it an environment feels like a bit of a reach. To me, a tool is something that changes your capacity, and these models absolutely do that. If I use a copilot to navigate a massive, legacy monolith, I am not being 'molded' by an environment; I am just finally able to understand a codebase that was written by three different teams over a decade. It is a utility. If you are worried about it malforming your thought process, maybe you are just not being opinionated enough with your prompts.

HostThat is a fair perspective. It is easy to get lost in the abstraction when the reality is just trying to get a feature shipped by Friday. Speaking of shipping, let us talk about developer blogging. A new survey of 181 developers looks at why people are using LLMs to ghostwrite their technical posts. The data suggests that for a lot of people, it is about overcoming the intimidation of a blank page, but interestingly, the vast majority of these writers are doing heavy editing or full rewrites afterward.

GuestHonestly, this is the most honest piece of tech journalism I have seen all week. Most people are terrified of writing because they think they need to be a 'thought leader' rather than just someone sharing what they learned. If an LLM helps a developer who has never written a word get their troubleshooting guide out into the world, that is a win for the community. I would rather have a hundred AI-assisted blog posts documenting obscure library bugs than ten perfectly written, gatekept essays.

HostYou are an optimist as always. I saw one commenter, hwayne, dismiss the whole discussion as off-topic for the site, which is pretty classic for that crowd.

GuestThey always say that! But look, the tool is only as good as the edit. If you use it to start the engine and then you drive the car yourself, you end up with something decent. If you just copy and paste the hallucination, well, that is on you.

HostI think that is a perfect place to leave it. Samantha, thanks for coming on and helping us parse through these threads. It is always a pleasure to get your take on these things.

GuestAny time, Daniel. I am heading off to try and break my own agentic loop for the third time today, so I am going to need all the coffee I can get.

HostGood luck with that. You heard it here first, folks, all of today's stories came from the tech community over at Lobsters. Thanks for listening, and we will be back with more tomorrow.