David Rose 🫧 (@drose101) 's Twitter Profile
David Rose 🫧

@drose101

Randomness and dynamical systems for fun, making you click ads for work. Digital nomad and ML engineer. Building AI agents!

ID: 194374752

linkhttp://drose.io calendar_today23-09-2010 23:55:56

840 Tweet

364 Followers

528 Following

David Rose 🫧 (@drose101) 's Twitter Profile Photo

> writes up a big spec after some research > puts together a plan with claude > 'work on this for a while im heading to bed' hit send > i wake up to this.

> writes up a big spec after some research
> puts together a plan with claude
> 'work on this for a while im heading to bed' hit send
> i wake up to this.
David Rose 🫧 (@drose101) 's Twitter Profile Photo

Spent an hour or so testing coding agents on plane wifi, Gemini and codex work great, Claude spends a minute per action, unusable . 🤷‍♂️

David Rose 🫧 (@drose101) 's Twitter Profile Photo

review your mcp tool schemas every once in a while i noticed my web search tool spent 4,000 tokens on timezone enums for every single prompt!

David Rose 🫧 (@drose101) 's Twitter Profile Photo

playwright mcp was like 13,000 tokens in full!, what a mess. re-wrote it in my own mcp wrapper, down to under 300 tokens as a single 'browser-hub' did the same for an openai web search tool any agent harness can call (gpt-5.2 web search is best) github.com/cipher982/mcp-…

David Rose 🫧 (@drose101) 's Twitter Profile Photo

Need to find a smoother mobile coding experience. Current setup is just using Screens on iOS basically a vnc style app but very polished. Quite a pain to zoom and pan around. Really just good for small follow ups in an ongoing session. Maybe it would work better on iPad

David Rose 🫧 (@drose101) 's Twitter Profile Photo

opus 4.5 loves to throw in a max_tokens on any llm calls it writes, which inevitably is too low (because it doesnt think about reasoning tokens) and crashes everything later on

denislav (@kuberdenis) 's Twitter Profile Photo

my girl just passed me and seeing my screen said "this is the most depressing shit i've seen.. 7 black windows.. all day long.. awful"

my girl just passed me and seeing my screen said "this is the most depressing shit i've seen.. 7 black windows.. all day long.. awful"
David Rose 🫧 (@drose101) 's Twitter Profile Photo

I wonder what the 'running multiple agents at once' flow looks like for most people. I have about 5-6 tabs on my terminal window, and just have a continuous loop of : review & respond, click tab to the right, review & respond, etc until i get to the end, then loop back to the

David Rose 🫧 (@drose101) 's Twitter Profile Photo

Thats a first, the new gpt 5.2 codex model just throwing in a random russian word? after i pasted a link to some google docs

Thats a first, the new gpt 5.2 codex model just throwing in a random russian word? after i pasted a link to some google docs
David Rose 🫧 (@drose101) 's Twitter Profile Photo

I've seen enough. tokens are cheap and lines of code are fast, im ripping langchain/langgraph out of all my projects. Almost all bugs in encounter result from fighting various abstractions from that ecosystem. i think i can rebuild bespoke implementations as needed for it all.

David Rose 🫧 (@drose101) 's Twitter Profile Photo

claude code when i post api keys into chat: 'thanks, i'll use those now' codex: 'oh no! you gave me sensitive information, rotate your keys and scrub this session'