Max Key

Tech lead for humans and AI agents • Closing SDLC feedback loop

[Max Key @etechlead]

@ajambrosino can't wait

[Max Key @etechlead]

@ZagZino @OpenAIDevs Which is especially insulting, considering that Windows is still the most used OS by developers

[Max Key @etechlead]

@bjornbjorn @unclebobmartin Kimi-K2.6 is good, yeah.

I've also heard good things from people using:
* Codex for planning
* Kimi, or another strong coding model, for implementation
* Codex again for verification

So I’d give that workflow a try, especially for backend

[Max Key @etechlead]

Yep, for me, Codex + GPT-5.5 High+ is far more rigorous when reasoning about a project, it can dive deep into complex areas, very rarely makes things up, and SOTA in working with context.

I review every major LLM release for SWE/architecture-related work, so this is based on my experience working with both CC+Opus 4.7 and Codex+GPT 5.5.

[Max Key @etechlead]

@deksden @simpsoka Да я просто каждый день наблюдаю эти мучения :)

[Max Key @etechlead]

@simpsoka * loosing track of subagents - they could disappear from the list or, worse, completely forgotten, especially between compactizations

[Max Key @etechlead]

* constantly using POSIX syntax in PowerShell - like "&&", which is not supported

* trying to use sed/grep/tail in PS (may be install them with Codex? :))

* incorrect symbol escaping and quoting in PS commands/scripts

* incorrect WSL paths: mixing up /c/Users/ with /mnt/c/Users/

* next.js + PS + [...] paths is a constant issue because [...] is special syntax in PS

* forgetting that windows executable file is locked when running - it creates problems to build and test executables

* loosing track of launched child processes - leaves multiple zombie processes which I have to kill manually

* sometimes Codex decides to kill zombies by itself and kills all the node.js'es on the machine, even non-relevant ones

* some old chats are disappearing from Codex App by themselves which is very annoying

* longer chats are slower and slower to render over time (it become better recently though)

* feature lag compared to Mac OS, sometimes implicit - for exampe, Pro, 26.506.31004 - no Chrome plugin (which marketed as available with no OS dependency)

[Max Key @etechlead]

@mohit_r9a @testingcatalog Thanks for the research.
Btw, there is nothing about GPT 5.5 in the paper.
Can we consider that xhigh reasoning was used for it as well?

[Max Key @etechlead]

@unclebobmartin Just to add some context - it's recorded on May 2025.
YEAR ago

[Max Key @etechlead]

@thsottiaux Windows support

[Max Key @etechlead]

@OpenAIDevs Bringing in the classic sheep.exe

https://t.co/yKEqQNxWbg

[Max Key @etechlead]

I couldn’t resist bringing the classic eSheep back from 1995, this time as a Codex pet.

Make your own here: https://t.co/8W9MY5bQQi
(it has a lot of other funny animations)

Or install the ready-made one:
npx codex-pets add esheep64

[Max Key @etechlead]

I couldn’t resist bringing the classic eSheep back from 1995, this time as a Codex pet.

Make your own here: https://t.co/8W9MY5bQQi
(it has a lot of other funny animations)

Or install the ready-made one:
npx codex-pets add esheep64

[OpenAI Developers @OpenAIDevs]

Pets. Now in Codex.

Use /pet to wake your pet.

[Max Key @etechlead]

@DragorWW На хабре довольно специфическая аудитория относительно использования ИИ в разработке :)

[Max Key @etechlead]

No product features were harmed in the making

[Max Key @etechlead]

@theo Couple more steps in this direction and we'll end up with SKILL.exe

[Max Key @etechlead]

@haider1 And some things have already been plateaued for a long time.
It's just ridiculous to believe in AGI

[Max Key @etechlead]

@bcherny @Rahatcodes It's like classic Bill Gates' code review "fuck counter" :)

https://t.co/PDAqnXTP2E