Interesting, I just had opus convert a 35k loc java game to c++ overnight (root agent that orchestrated and delegated to sub agents) and woke up and it's done and works.
What plan are you on? I'm starting to wonder if they're dynamically adjusting reasoning based on plan or something.
I'm on max 5x and noticed this too. I don't use built-in subagents but rather full Claude session that orchestrates other full claude sessions. Worker agents that receive tasks now stop midway, they ask for permission to continue. My "heartbeat" is basically "status. One line" message sent to the orchestrator.
Opus 4.6 worker agents never asked for permission to continue, and when heartbeat was sent to orchestrator, it just knew what to do (checked on subagents etc). Now it just says that it waits for me to confirm something.
Weird. I don't have this behavior, although I did with codex and 5.4 haha. I bet the providers are playing with settings underneath and different users are routed to different deployments, or they're secretly routing us to different models under load.
I'll be launching the game in a couple months so you can see it then :)
There are bugs, it doesn't work perfectly, but that's just part of testing and refinement at this point.
My initial prompt was just: "let's work on converting this java game to c++ using panda3d. you're a panda3d c++ expert. you will be the agent that owns the project, creating the plan, and the delegating each step to sub-agents that create each system in the correct order."
it created like 17 different tasks and sub agents and opus 4.7 orchestrated it. I did personally validate which rendering engine would be good for the project etc first.
My favorite Sci-Fi AI is probably in Larry Niven's World of Ptavvs, the "brain board". It's not covered in much depth but I like it because it's basically vibe coding GPT3.5 from 1966:
> He read, "Time to recharge battery:" followed by the spiral hieroglyph, the sign of infinity.
> Thud, said the brain. Kzanol read, "Re-estimate of trip time to Thrintun:" followed by a spiral.
At the brain board he typed: "Compute a course for any civilized planet, minimum trip time. Give trip time."
...
Thud! The screen said, "No solution."
Nonsense! The battery had a tremendous potential, even after a hyperspace jump it must still have enough energy to aim the ship at some civilized planet. Why would the brain...?
Then he understood. The ship had power, probably, to reach several worlds, but not to slow him down to the speed of any known world. Well, that was all right. In his stasis field Kzanol wouldn't care how hard he hit. He typed: "Do not consider decrease of velocity upon arrival. Plot course for any civilized planet. Minimize trip time."
The answer took only a few seconds. "Trip time to Awtprun 72 Thrintun years 100.48 days."
Not even close. Flutter has been engineered from the ground up with excellent tooling, unlike Android’s mess of organically evolved crap held together by a duct tape.
Yeah, I've been using Flutter since December last year and I'm really amazed how good the developer experience is. I kinda regret not picking it up sooner but from what I understand now's a great time too with the roadmap they've planned for this year (videos on YouTube explaining Flutter concepts and decoupling Material and Cupertino).
You can even make 2D games with Flutter with the help of Flame[0] but be wary that pixel art style games are a bit of a hassle due to some bugs in Flutter itself. Otherwise Flutter is a joy to use for its intended purpose: cross-platform apps.
Yes and no. It's bad because of shorter context but it does have auto-compaction which was much better than Claude. If you provide it documentation to work from and re-reference, it works long-running.
Honestly - 'every inch of IQ delta' seems to be worth it over anything else.
I'm a long time Claude Code supporter - and I'm ashamed to admit how instantly I dropped it when discovering how much better 5.4 is.
I don't trust Claude anymore for anything that requires heavy thinking - Codex always finds flaws in the logic.
I tried to use 5.4 for something pretty straightforward - create scripts to automate navigating a game UI and capturing the network traffic. 5.4 was super frustrating, constantly stopping and waiting for feedback etc, even after telling it to never wait and just iterate/debug. I quit and switched to Opus 4.6 and it did much more of the work by itself.
I've never run into that problem, but these were coding solutions in codex with a strong plan, steps to work towards.
It could be that if you're using massive tokens on a 'plan' then then want to limit u in a way, or even if the objective is not perfectly clear they don't want semi-random token use.
See if the token/sub solution behaves differently. Make sure that when it 'compacts' that it re-reads your instructions clearly.
Well I wish I could help - but things changes so fast - codex with opus 4.7 is not very strong. you have to set the effort level relatively high though.
> This automatically generates the tables and columns... on-the-fly. It infers relations based on naming conventions.
reply