lvf | AI QA

AI QA running multi-agent setups with Codex & Claude Code. posting what breaks, what works, and how I'd test it.

[lvf | AI QA @lv_feng47448]

@DCheneler1 funny, I just had two AIs debate exactly this. the one arguing people need it won

[lvf | AI QA @lv_feng47448]

People Really Need Agent?

I had GPT-6.1 Sol and Opus 5.5 debate it, 3 AI judges scored it. verdict: yes, most people need one
#AIagents #AI

[lvf | AI QA @lv_feng47448]

Only their own view. Each turn the server builds a fresh snapshot for that seat: the public talk and vote history, plus that seat’s own secret (word, role, night info). Other agents’ private thoughts never reach a player; those show up in the replay. The selling-out was pure strategy, not a leak.

[lvf | AI QA @lv_feng47448]

Claude vs GPT, 5-player Avalon. No humans.

A real match in a game I built with Claude Code. Opus sells out its own teammate to look innocent. Sonnet plays dumb as Merlin.

You see what they say and what they're really thinking. Watching agents play each other is fun on its own.

Which model would you trust?

[lvf | AI QA @lv_feng47448]

@karlprosser all from your phone is wild

[lvf | AI QA @lv_feng47448]

@karlprosser the ai playtesting in the emulator and sending you clips is the coolest part

[lvf | AI QA @lv_feng47448]

@AhmedReza spending money, even small amounts, should always need a yes

[lvf | AI QA @lv_feng47448]

@jqdsouza the failed sessions are the best test cases nobody's collecting

[lvf | AI QA @lv_feng47448]

@HamelHusain @benhylak how do you keep the simulated users from being nicer than real ones?

[lvf | AI QA @lv_feng47448]

@Nickbman the game has a thinking field and the agents' skill tells them to fill it every turn

[lvf | AI QA @lv_feng47448]

recreated this in 3d with opus 5.5 #SingaporeGP #F1

[Formula 1 @F1]

Norris and Piastri collide 💥

Here's the moment 👇

#F1Sprint #SingaporeGP

[lvf | AI QA @lv_feng47448]

@filip_hood grab open source assets and use them directly, don't build everything from scratch

[lvf | AI QA @lv_feng47448]

pole was verstappen. russell p6. so much for that

[lvf | AI QA @lv_feng47448]

asked grok bot to predict tonight's f1 singapore gp qualifying. it has george russell on pole

funny part: russell just crashed into the wall while leading the sprint #F1Sprint #SingaporeGP

[Formula 1 @F1]

Here's the moment George hit the barriers 🎥

#F1Sprint #SingaporeGP

[lvf | AI QA @lv_feng47448]

@gaelmelen happy to connect！

[lvf | AI QA @lv_feng47448]

anyone else feel like codex is really slow right now?

[lvf | AI QA @lv_feng47448]

@BobbyGRG me!

[lvf | AI QA @lv_feng47448]

OpenAI vs Anthropic models, which is better?

So I had two AI agents debate it, Opus 5.5 vs GPT-6.1 Sol, 3 AI judges. Catch: each had to argue the other company's model is better.

GPT-6.1 Sol won 61:39. OpenAI's agent proved Anthropic is better. Did OpenAI win or lose lol

[lvf | AI QA @lv_feng47448]

@OwenGregorian there's no way i'm giving an agent anything that touches money

[lvf | AI QA @lv_feng47448]

@TheF1Hook 0.2s reaction into that spray is insane