decided to share my recent observations from working with AI coding agents / LLMs everyday (subjective option):

• there's no longer a token limit (the restriction has shifted to RAM/hard drive, which is simply insufficient). Codex, Cursor, and other agentic harness builds / CLI tools - have solved the problem of context and LLMs in the loop, each in their own way, but roughly similar. Zed gave ability / glips how to work in realtime for humans, but git systems unfortunately doesn’t support realtime collaborative editing / coding (imho).
• the next important limitation is the chat direction vector, which is set in the first sent message.
• all frontier models are roughly similar, but with their own features.
• everyone's agentic harness is different, with the best being Codex (computer use (as a partner - full computer control, automation, + skills + orchestration) and Cursor for augmentation (multitasking & almost live editing & need for speed + orchestration). Claude code / design - feels like experienced but unpredictable person - ideal for research / discover new ideas and products / design, but requires constant monitoring.
• agentic harness is the current bottleneck, because it needs to be assembled manually. Ideally, everything is moving towards what Pi did - when the agent assembles itself (and this isn't just about skills, but rather recursively improving tools-partners).

_image for gradual transformation from drawing to realistic _

in my opinion, there's a sense that a new product and a new operating system will emerge within which the next round of development will take place (as seen with mobile phones - Blackberry / Symbian -> iOS / Android (I don't include Windows/Linux - there wasn't mass adoption) -> Aluminium OS / Google Books? with local models?).

_image for gradual transformation from drawing to realistic _

I think mass adoption will happen precisely when locally run models can do the same as remote models (as it was shown in Google IO with audio glasses) - because why need a browser and apps in OS in which developers needs to embed an agent, if you can install an agent (harness + llm) instead of an app and a browser, and it will already draw the interface and show / help manage the data accordingly to provided interface guidelines, design systems and user flows (it feels like design - in Excel VBA we’ve designed boxes with strict sizes, but since web we started defining constraints instead of fixed - so it may happen with how agents represent or manage information - we will define constraints (design / data / tool systems) - agents will adapt it for user)

It's also very interesting what will be possible in terms of creative content - games, assets, etc. Because AI is not just generation - the more I worked with models, the more possible to do, it feels just like 3D/ digitalization moment - instead of drawing everything, we can define 3D models and then do. iterations with it

_image for gradual transformation from drawing to realistic _

disclaimer:
image cover is generated by ChatGPT image 2 in Leonardo
images in text - grok imagine
text is autotranslated from my manual writing and rewritten manually again with many more details:)