in conversation with machines since 2015. now they talk back. → little tools for daily agentic work → mostly for fun, sometimes useful founder · builder
Joined March 2026
- Tweets891
- Following494
- Followers78
- Likes395
Pinned Tweet
I know there are already great coding-agent session managers and agent-aware terminals like Herdr.
But I’m so deeply invested in tmux that switching would mean rebuilding the workflow I use every day. I know tmux has limitations, but it still works incredibly well for me.
So I built tmux-agent, a small local-first tool that helps me monitor and manage coding agents running in tmux, regular terminals, and across multiple machines over SSH and Mosh.
It supports Codex, Claude Code, OpenCode, Grok, OMP, and Pi.
I built it primarily for myself and I’ve tried to keep it as workflow-agnostic as possible: it doesn’t replace your agents or force a particular orchestration system. It simply adds visibility to the terminal workflow you already have.
If you’re also deep into tmux, maybe it can help with your workflow 👇
GPT-6 Astra is the first model that feels like it can hold a 3D scene while you keep talking.
Fable 5.1 was already a leap. Astra is more plastic. more control over the environment.
robotics stunned me the same way. it opens new ways to train and work with robots. human input to action. look, miss, adjust.
you do not plug the model in and watch it walk. you still need the hardware. you still need the principles. people who already know that stack just got a much faster loop.
pure acceleration. the door opened. the people who already had the loop are running through it.
i feel like we're all getting used to this rapid spike in intelligence after each model release.
it's still inspiring, but i don't get goosebumps anymore.
ok my investigation is progressing. i tried almost everything and i think this is not related to CLIProxyAPI. for now my best bet is that the OpenAI account is busted. i tried the same prompts on a fresh setup. fresh Codex install, no skills, no plugins, nothing, just bare Codex. same issue.
some runs look like the agent is dropping reasoning and just delivering the first attempt. it's not iterating, not checking work. or it looks like it's running with almost no thinking. reasoning tokens vs healthy runs are something like 500 vs 10,000.
i tested the same prompts on another account and the results are always consistent. high quality. with the account that causes the problem it's always hit or miss. sometimes ok-ish, sometimes ultra bad.
I’ve been using CLIProxyAPI since February or March. It’s a great and very convenient piece of software.
However, lately I’ve been experiencing issues, especially when using GPT-6 Astra. The model often seems to stop reasoning much sooner and puts far less effort into completing the task. It’s particularly noticeable with 3D modelling, graphics, and other visual work. It may also affect coding, but that’s sometimes harder to spot.
I gave Astra exactly the same task through CPA and through native Codex OAuth. The difference was night and day.
Native Codex recreated a 3D scene from a reference image almost perfectly: lighting, mood, textures, composition, everything. The same task through CPA produced something that looked like it had been made by a five-year-old in Paint.
I’ve done a lot of testing but still can’t figure out what’s happening. It doesn’t affect every run, and CPA occasionally performs well, but most runs are noticeably worse.
The difference is clearly visible in the reasoning traces and token usage. Native Codex works for 10–20 minutes, iterates repeatedly, checks its own output, and delivers excellent results. Through CPA, it often stops after 3–5 minutes, makes far fewer tool calls and checks, and may use around 500 reasoning tokens instead of 10,000.
It’s extremely frustrating. Is anyone else experiencing the same issue?
I’ve been using CLIProxyAPI since February or March. It’s a great and very convenient piece of software.
However, lately I’ve been experiencing issues, especially when using GPT-6 Astra. The model often seems to stop reasoning much sooner and puts far less effort into completing the task. It’s particularly noticeable with 3D modelling, graphics, and other visual work. It may also affect coding, but that’s sometimes harder to spot.
I gave Astra exactly the same task through CPA and through native Codex OAuth. The difference was night and day.
Native Codex recreated a 3D scene from a reference image almost perfectly: lighting, mood, textures, composition, everything. The same task through CPA produced something that looked like it had been made by a five-year-old in Paint.
I’ve done a lot of testing but still can’t figure out what’s happening. It doesn’t affect every run, and CPA occasionally performs well, but most runs are noticeably worse.
The difference is clearly visible in the reasoning traces and token usage. Native Codex works for 10–20 minutes, iterates repeatedly, checks its own output, and delivers excellent results. Through CPA, it often stops after 3–5 minutes, makes far fewer tool calls and checks, and may use around 500 reasoning tokens instead of 10,000.
It’s extremely frustrating. Is anyone else experiencing the same issue?
ok i'm replacing Muse Spark 1.2 with GLM-5.3-Flash.
it's amazing, and none of that feeling that you're running a "less intelligent" model. there are moments i'm genuinely surprised by the output.
Muse is great. very solid. GLM-5.3-Flash is the one that actually made me interested.
it's like with girls. some are just nice and pretty. a few are interesting. that's the vibe this one has.
landing in my daily stack.
February was the @openclaw moment.
It changed how many of us thought about AI. Agents stopped being demos trapped inside browser tabs. They became persistent, personal, capable of doing real work through the tools and conversations we already use.
Was OpenClaw ready for everyone? Probably not. It was rough around the edges. Things broke. You had to tinker, debug, repair, and occasionally question your life choices.
But that moment was necessary.
Six months later, @bot arrives with a very familiar idea, packaged into something more polished and easier to use out of the box.
That doesn’t diminish what OpenClaw started. It validates it.
I still use OpenClaw every day, and it works incredibly well for me. I want the control. I enjoy experimenting, building, breaking, and rebuilding.
Others just want to press a button and get a working agent. Grok Bot will be the better product for them.
OpenClaw and Hermes from @NousResearch will remain homes for people who want to own, shape, and understand their agents.
This is how technology matures. First the breakthrough, then the productization, then the market decides.
There is room for all of them.
to understand how qwen 3.8 27b thinks is the key to understanding how to use this model effectively.
tested the latest Dynamic V3.0 by @UnslothAI.
I wanted to see exactly how Qwen 3.8 behaves differently with Low and X-high reasoning, so I dug through the raw traces.
Qwen 3.8 27B Dynamic V3
UD-Q6_K · 256K · KV4 · MTP
➡ Across 3 × 50 coding-agent runs, Low and X-high scored the same on easy and medium tasks.
X-high only paid off on the hard tasks.
New Unsloth Dynamic V3 GGUF is a real improvement for Qwen 3.8 27B.
🏁
UD-Q6_K,
llama.cpp,
256K context,
Q4 KV,
MTP draft max 2,
full GPU offload on RTX 5090.
Across my agent runs:
➡ 108 tok/s average decode
➡ 1,074 tok/s average prompt processing
Fast and sharp.
testing new Qwen3.8-27B GGUFs 👀
We’re releasing new Qwen3.8-27B GGUFs with 10% higher accuracy.
Unsloth Dynamic V3 outperforms others by >10% on Div-300, KLD & more benchmarks.
We also release 1-bit quants that retain 77% accuracy. Run on 8GB RAM.
Blog: unsloth.ai/docs/basics/dynam…
GGUF: huggingface.co/unsloth/Qwen3…
The thought that keeps me awake at night:
What if I recorded my entire life?
Not just photos, videos, or the moments worth posting. I mean everything I could capture: every conversation, every thought, every belief I changed, every fear, mistake, realization, and ordinary moment - transcribed into one Markdown file per day, from now until the day I die.
If a future neural network could read all of it - not only facts about me, but decades of how I speak, react, doubt, love, decide, and evolve; could it simulate me?
Could my children or grandchildren talk to that version of me after I’m gone? Would it remember the same people, tell the same stories, make the same jokes, and reach the same conclusions I would?
And if the simulation became accurate enough, what exactly would survive?
Me?
A copy of me?
Or only an extraordinarily convincing echo built from everything I left behind?
I don’t know whether that would mean I never truly died, or whether it would simply be the most detailed ghost humanity has ever made.
But I can’t stop thinking about where the boundary is between preserving a life and preserving a person.