Eric Provencher (@pvncher) just dropped a really solid write-up on rethinking skills, AGENTS.md, and prompts now that GPT-6 Astra is out. Short version: a lot of the scaffolding we piled on over the last year is now working against us. Bloated skill descriptions, "read the whole repo first" habits, over-specified recipes, and boundaries that make Astra stop too early. Worth a house-cleaning pass. Go read it: https://x.com/pvncher/status/2095991462416490862

I’ve been running pstack inside Cursor for a while. It's a great setup with one problem. Every delegate runs at Cursor usage rates, so the models I actually want on hard tasks were too expensive to use by default and maxing out my free other usage that I prefer go to bug bot. Meanwhile I'm already paying for Claude Max, ChatGPT Pro, and SuperGrok Heavy. I wanted the pstack playbooks with my subscriptions doing the work. So I forked it. Grok Bot became the orchestrator and T3 Code became the execution layer which would give me similar workflows to using grok bot with cursor cloud agents except the cloud is my hardware, and I can use my subscriptions instead of cursors api rates for my other models as well as incorporating my local models and open router if necessary. And I can do it all from my phone which allows me to leave my office and still get work done :) Grok Bot holds the router, the playbooks, and the table that says which model handles which kind of job. It decides what runs where and writes the brief. T3 Code runs on a Mac Studio in my home network and wraps the coding CLIs I already pay for: Claude Code, Codex, Grok Build, and a local model. For each step, Grok Bot opens a T3 thread on the right provider, sends the brief, waits, and reads the results. My original plan was to have Grok Bot use T3 Code the way I do: through the app. That went badly. T3 is built for a human with a screen, and the Bot struggled to drive it reliably. It could get a thread open, but sending work in, knowing when the delegate was actually finished, and pulling the result back out was fragile every time. An orchestrator that can't tell "done" from "still thinking" isn't an orchestrator. The fix was to stop asking the Bot to use a UI and give it a tool shaped for a bot. We built a small command-line tool with exactly the handful of actions the playbooks need: start a thread, send it work, wait for it to finish, read what came back, cancel it. Every action has a clean start and a clean end. Once the Bot had that, the pilot lanes ran hands-off. That's the biggest lesson in the whole project: if an agent is fumbling a tool, don't write a better prompt, build a better interface.

It quite surprised me that it's IBM doing this :) For anyone lost in the sea of terminologies, principles and needs guidance (like me), I recommend the channel bellow. IBM tutors go through different subjects in a clear and short way. https://youtube.com/@ibmtechnology?si=Mfq1iP1VVHlTf8b2

I’ve been running open-weight models locally for a while. GLM-5.2 on a 512GB Mac Studio was capable, but you always felt the ceiling: 2-bit quant, 5-9 tok/s, tool calls that got flaky in long agent sessions. Local was the thing you tolerated to keep data on your network. DeepSeek V4 Flash was the first real step forward — finally a local model with a genuine balance of performance and speed, one you could actually leave running under agent traffic without babysitting it. It proved the “fast MoE with real agentic chops” formula worked. 5.3-Flash on two DGX Sparks takes that further. 320B params but only 18B active, so it’s running at good speeds. 1M context. Thinking is always on but tunable (low/high/max), and at max it holds up in real agent work through Hermes: multi-step coding, sustained tool calls, no hand-holding. I’m still reaching for cloud models. But for the first time I can see a point coming where you won’t have to. The gap is closing faster than I expected — if you have the hardware, this is the one to try.

https://youtube.com/live/AAw5r0e-ozM?feature=share