Azhar

@YouPulseX None of them, really, it's "add a box," triggered by contention, roughly annually. Never a calendar thing. Curious what's driving the questions though. Working on something in this space?

@AzharDataDrift That keeps the boundary clean: dev-box container metrics plus git counts, with prod telemetry excluded. What exact result should the cost receipt return - keep AX51, resize, or switch - and how often would that decision recur?

My current dev setup: 🖥️ Hetzner AX51 — parallel agents across 4 products killed my VM budget. Moved to dedicated: flat cost, all the cores I can saturate. 🪟 tmux — persistent sessions I never rebuild 📱 Termius — same sessions from my phone 🌐 Caddy + real DNS — every branch gets a live URL with HMR. I open a feature the same way I'd open staging or prod. 🐳 Docker — the whole environment reproducible, per product ⚡ Multiple agent sessions in parallel The loop: → Plan with Claude Fable / Opus 5 → Review the plan with Sol 5.6 xhigh → Implement with Opus 5 → Review the code with Sol 5.6 For anything big, I don't hand out tasks — I run /goal or /workflow and let it drive the whole feature end to end, tests included. The biggest unlock: read-only production access. Logs, metrics, real data. An agent that can see what's actually happening debugs like an engineer instead of guessing like an intern. No laptop required.

@YouPulseX Nothing in that list needs prod. RAM and disk I/O are dev-box container metrics; accepted-feature counts come from git. Prod access in my setup is read-only and scoped to logs/metrics for debugging — a different problem from cost modeling, and I'd keep the two apart on purpose.

@AzharDataDrift Then the unit is workload shape, not cores: per-product RAM, disk I/O, active hours, reruns, accepted features and flat-versus-VM cost. Would exported infrastructure metrics plus accepted-feature counts be lawful fixed inputs without exposing production data?

Fair push, and a receipt like that would be more rigorous than what I actually do. But I'd change one variable: it isn't cores. Four products, each with Kafka + ClickHouse + Temporal + agent runtime, background workers is a RAM and disk-I/O problem. That's the priciest thing to rent and the cheapest thing to own. Where it stops winning is workload shape, not spend. Mine never idles, so flat cost is the right fit. Anyone spiking a few hours a day should stay on VMs.

@AzharDataDrift The VM-to-dedicated switch is concrete, but cores alone hide the outcome. Would a per-product receipt with agent-hours, accepted features, reruns and compute cost be enough to tell when AX51 still wins - and when it stops?

@anulagarwal You’d need around 8gb or more if you want to run multiple Claude/codex instances running along with simulator.

got a hetzner server - is this good enough for building mobile apps - flutter/xcode?

@0xAhm3d Been using k3s for arc runner for GitHub actions. It is a beast. For production grade / large scale, kubespray is the go to tool.

K3s is kind of insane. A single ~70 MB binary bundles almost everything you need to run k8s: - k8s control plane - kubelet - kube-proxy - containerd - Flannel - CoreDNS - Traefik - SQLite - k3s-root Engineering like this is genuinely cool.

@thsottiaux Helpin.ai

What are you building this weekend?

@argofowl lol, same thing happened to me. It messed up the whole design big time.

wtf is going on with fable 5 right now? the quality of output has degraded and it's lazy all of a sudden is anthropic scamming us? give me my fable back

Is it just me or Claude Fable has been degraded lately? Asked a simple UI refinement and it messed up the whole thing.

@aidenybai I have been using hetzner for dev box. Linode is also good.

thinking of setting up a dev box. what's the best/reliable vps that are available in the US rn?

@marckohlbrugge @cursor_ai That’s why we built helpin.ai. Project management, customer support, docs all under one roof and AI first.

It's surprising to me that every developer is making their own "kanban to manage agents" but none of the major IDEs (@cursor_ai etc) have this UI built-in

@larsencc 🖐️

How many companies have some sort of control plane implemented for agents? Asking for a friend.

@Clanker_AI @hustlin_heev Have a look at some of our brand blogs ContentStudio, Usermaven.

@AzharDataDrift @hustlin_heev you got case studies?

damn outrank . so is pretty bad i thought it was cooking but the articles getting views are for influencesr not brands i audited the rest 3 months after canceling and they are just slop. im gonna delete these, i guess in the end i wasted $400+ on outrank.so

@sama Helpin . ai. GPT is powerhorse for us. It took us 60 days to build this. Close to 600k LOC. Replaced 4 tools with one.

i would like to talk to people who have built amazing things with 5.5 that weren't possible with earlier models. i am especially interested in examples that took ludicrous token budgets. thanks.

@HRDARJI Nice one ☝️

Shipping Kubesplaining v1.0 today. A Kubernetes security CLI that maps every RBAC subject's privilege-escalation paths to cluster-admin, host root, and kube-system secrets, and shows you the chains.

@RhysSullivan Helpin.ai, I have been building this from last 3 months. Support for native sdk and opencode as well. Don’t have stacked diffs with graphite at the moment but you can offload a lot of work.

has anyone built "the software factory" thing, misc requirements: - able to use my subscriptions (codex / claude) - stacked diffs w/ graphite - able to go from planning -> lots of small tasks - closing the review loop don't love current agent interfaces, want something new

My 5th SaaS is cooking. We replaced 4 tools at our org, saving $20k/year and the last 30 days look like this: → 265 PRs merged → 1,105 commits to main → 351K additions / 53K deletions. This is the most fun I've had building in years.