Monter

Building AI products solo. I’ll post what I’m working on, what I’m thinking about AI and tech in general, and where I think things are going. Mostly for myself, but public. 🧠

@WVVWVVWVWVVV @aryanlabde Mostly OpenClaw for orchestration, then IDE/CLI when it makes sense. For me the useful bit is not the surface. It is routing the work and checking /status so I know what model is actually doing the job.

@buildwithgoks @aryanlabde Are you guys using codex app or IDE extension

As a vibe coder, which plan gives you the best value right now? - $20 claude - $100 codex - $200 claude

@sflorimm Both 🚀

Are you paying for Codex or Claude or both?

Yea setup was the painful part tbh. I started with the manual workflow first: what tickets come in, who owns what, what context they need, what can go out without review. Then I turned that into files + rules the agents read every run. If the human workflow is messy, the agents just automate the mess.

@buildwithgoks @TTrimoreau Love this classic clarification man. Agent of course does a lot of work that we as human developers back then can only think about. But to be honest it is the setting up that sucks the most. How did you go about that?

Does anyone actually run more than 1 agent in parallel?

@AllenTheDetails @aryanlabde For claude on openclaw, you can run it via the claude-cli provider

@buildwithgoks @aryanlabde What’s your workflow for routing? Are you using Claude Code in terminal?

@pcshipp For frontend specifically I’d probably test GPT 5.5 first right now. Claude still feels better when I need the agent to reason through messy product context, but repo/ui iteration is where Codex has been getting harder to ignore for me.

Which LLM is best for frontend development right now? - GPT 5.5 - Opus 4.7 - Gemini 3.1 pro

@AllenTheDetails @aryanlabde Not Claude Code in terminal for the main flow. I use OpenClaw as the orchestrator and route by task/model from there. Repo-heavy stuff goes Codex, reasoning/personality goes Claude, cheap models for boring repeats. Main thing is checking /status so one sub doesnt eat the month.

@duborges I don't trust those 18 hour runs either unless there are checkpoints. Long running agents can work, but if it disappears for half a day and comes back with a monster diff, that's not autonomy to me. That's stress.

i think most tweets about Codex/Claude saying: “i let my model run for 18 hours straight…” are bs coming from social media maxxxers i’ve never had any tasks taking longer than 1 hour, and if it ever happens, i would be really concerned about the results after so many tokens spent, there’s no way the model would still perform well without cleaning sessions one hallucination could become a huge pile of shit in the end but i could be wrong and there’s this incredible framework or coding style that i’m missing?

@witcheer Depends what you want local for. If it is coding + agents, RAM matters more than people admit. My 16GB Mac mini already feels tight, so I’d start with what models you actually want to run before buying anything. Local is sweet until the hardware becomes the bottleneck.

IT IS TIME I want to switch to local AI. I asked Grok and Claude what should I buy and read according to my current set up. I would love to get any advice on these answers, if it is coherent and relevant or if you would advise something different as a starting point. I have $1k - $2k to spend my current hardwares are: 1. Mac Mini M4 (16GB RAM) 2. Windows desktop tower: > CPU: AMD Ryzen 5 7600X (6 cores, AM5) > Cooler: Arctic Liquid Freezer II 360 A-RGB (360mm AIO) + Arctic MX-4 paste > Motherboard: ASUS TUF Gaming A620M-PLUS WIFI (mATX, A620 chipset) > RAM: 32GB DDR5 Corsair Vengeance (2×16GB) 6000MHz CL36 > GPU: MSI GeForce RTX 4060 Ti VENTUS 3X OC (8GB) > Storage: Kingston KC3000 1TB NVMe > PSU: MSI MAG A650GL (650W, 80+ Gold) if you know your shit I would greatly appreciate feedbacks and advice.

@pcshipp First thing I’d check is whether they understood the value before pricing. 1k active users and 0 premium usually means onboarding/positioning is broken, or the paywall is showing up at the wrong moment. I’d talk to 10 users before touching the price.

You have 1,000 active users but not a single user purchasing premium What’s your next move?

@PenuelAkom @TTrimoreau Mostly boring setup, not magic. Each agent has: - a clear scope - memory/context files - tools it can and cant use - handoff rules Then Orion routes the work and reviews outputs before anything visible goes out. The constraints matter more than the prompts tbh

@buildwithgoks @TTrimoreau This must be doing a great job. How did you set this up?

Anyone using Gemma 4 local model with openclaw? Whats the performance like for you 🤔

@big_duca Yea. Daily. Mostly for splitting boring operational work across agents: engineering tickets, social, security checks, admin stuff. It’s not magic out of the box though. The value only starts showing after you treat prompts/skills/memory like actual operating docs.

Is anyone still using OpenClaw?

@PenuelAkom @TTrimoreau I keep it pretty boring: - one owner per task - narrow scope - shared memory/context files - one place to review outputs The manager agent mostly does routing + follow ups, not the actual work. Otherwise you just get 5 smart agents making 5 different messes.

@buildwithgoks @TTrimoreau Real man. Can you send me your implementation approach?

The only way I’d see that happening is if the agent is running different models because they are effectively different sessions and I’ve noticed sometimes Openclaw can internally switch models without your Knowledge, so you have to either run /status or ask it what model it’s using to confirm

Has anyone perceived different performance in OpenClaw by the same agent on the same model but on a different communications channel (e.g., TUI vs. Signal vs. Matrix)? Because I feel like on one of those, an agent is even smarter than usual. Which shouldn’t be possible, at least not strictly on the basis of channel selection. Hmm…

@bradmillscan @steipete @cathrynlavery Also adding a LESSONS.md file where you can add all the things it should stop doing and it should load it into context

@bradmillscan @steipete @cathrynlavery I don't think you fully make it forget by shouting at it. What has worked better for me is cleaning the files it keeps reading every session. AGENTS.md, skills, memory, old weird rules. Treat it like onboarding a new employee after the company process changed.

Is there a way for your OpenClaw agent to "forget" all of it's bad habits & relearn how to do things based on new OpenClaw standards? It doesn't make sense that starting over is the only way to get your agent to stop doing weird Feb/March quirks. cc @steipete @cathrynlavery

@indiesoftwaredv This is why I don't dismiss boring formats. Everyone wants the fancy AI video, but if a 5 min slideshow gets similar reach, the math is the math. Distribution humbles you fast lol

Receipts from yesterday's claim 👇 My TikTok grid this week: Slideshow → 1,144 views ✅ Slideshow → 575 views ✅ Slideshow → 410 views ✅ AI video → 574 views ❌ AI video → 507 views ❌ AI video → 448 views ❌ 5 min vs 1 hour to create I'm done with Kling (AI Video) for now

@LysandreJik This is the boring bit that actually matters. Agents can produce something once. The harder part is knowing it still works after the docs, examples and edge cases change. Benchmarks for agent workflows are going to be underrated imo

I've been trying to make transformers more agent-friendly: agentic CLI, a skill, doc rewrites, canonical examples. It felt a bit like shooting in the dark: hard to measure progress, and hard to ensure what worked once will continue working. So we built a benchmark!

This is the part people skip. The app is one asset. The distribution system around it is the actual machine. 10 apps without repeatable acquisition is just 10 jobs lol x.com/alexcooldev/st…