Agentic Lab

The creator of the Ralph Wiggum Loop @GeoffreyHuntley just declared my video on the loop as the official explainer video. Here's the run-down and core ideas (see thread)

@bradmillscan Bro please lock in and just build your own at this point

350-400 hours into OpenClaw over the last 33 days non-stop, no days off...I'm ready to quit. My openclaw is fucking lost in the weeds every day today and it's driving me nuts. Basic shit. I asked it to use GitHub. it has a GitHub skill. We have a GitHub SOP. I can see it's thinking process about using skills, then narrating how the skill doesn't exist, then going and inventing ways to retrieve the capability to use GitHub from the internet. I tell it to look in the openclaw docs for the proper skill path, it says "oops my bad, yeah it was there after all." This is ChatGPT 5.4 with extra high thinking turned on. I ask it to diagnose the problem only, so it goes and sees the system prompt is telling it to look at the wrong place, and it goes to GitHub and opens a GitHub issue about this 'bug' without even asking me. What the actual fuck. 3 hours on a Sunday of trying to rewire the brain of my openclaw to do default-behaviour. This thing such a productivity suck & mental poison. I can't do anything useful or positive with OpenClaw because I'm nonstop fighting fires in the engine room. I'm thinking about giving up.

@dexhorthy If NIAH is rotting this hard, think of what happens with coding in complex repositories 😭

don't make me tap the sign

As someone who was against AI for a long time, AI is a skill just like any other, and the usefulness of AI increases as you get better at it. The first time you play a guitar, you literally can't do anything. You sound like shit. But you slowly get better. You can play more songs. You enjoy it so much more. We were made to think that AI would just solve all our problems and do everything for us, so people were disappointed when it didn't do those things, but if you actually try to learn it, then you will see how helpful it can be. That sounds like a cop out, but it's silly to think that AI is just going to solve all your problems out of the box. If you don't use it for an extended period of time then (as with every other skill) you will not be able to get the best results with it. It won't remove all of your work, or any of it for that matter if you don't want it to (I've been spending more time working, because it's fun), but it will give you a deeper, more enjoyable way of doing what you already do. You abstract out a layer from physical labor to mental labor. You spend more time in the realm of ideas and taste. It starts to feel more like a game or puzzle than it does "work." You have to be okay with moving from doing the manual grunt work of writing lines of code or prose to discerning whether or not it sounds like you want it to sound. I don't think anyone, aside from other writers who haven't detached from the labor, would argue that hiring a ghostwriter (i.e. James Patterson) is cheating. The end readers still love it. He is still in control of the plot, vision, and what actually gets published. Nobody would argue that the director of a film doesn't deserve praise because he is not manually filming it or building the set. There is going to be, and there already is, so much pushback on AI for valid reasons, but most of those reasons are psychological defense mechanisms tying you to how you've identified yourself with how you work. AI doesn't have to be a religion. You don't have to use it for everything. But I don't think it's a wise idea to willingly choose a side that is going to lose.

@TukiFromKL PSA: Code Review is literally just a prompt. They package it as a feature to stir up hype.

🚨They're not even pretending anymore. > Claude just announced that when you open a pull request, it sends a TEAM of AI agents to review your code. > The thing your senior developer charges $200K/year to do? An AI army now does it the second you hit submit. > And they called it "dispatching agents to hunt for bugs." Hunt. Like your codebase is prey. > This morning Microsoft launched an AI that does your entire office job. > Tonight Anthropic launched an AI that does your entire engineering job. Two companies both came for your salary. Tomorrow's Tuesday and I'm honestly scared to check what drops next. 💀

@kissapiai Yes, it gets even more effective as we control context better!

@AgenticLab1 that's what makes it effective tho. not trying to be AGI, just giving the LLM hands and eyes. the magic is how simple each piece is but how powerful they get when composed. plus you can swap models - Opus, Sonnet, GPT-5, whatever fits the task

Everyone's calling Openclaw AGI. I reverse-engineered the entire architecture. It's a set of building blocks around an LLM. Heartbeats, cron jobs, JSONL persistence, tool schemas, and browser control. An exoskeleton that gives the model the capacity to perform complex tasks on a computer. Impressive engineering. But after a month of daily use, you're looking at ~45,000 tokens of fixed overhead before you even send a message. That's up to a 40% performance decrease from context rot alone. A single-purpose agent needs ~1,400 tokens of overhead. Openclaw needs 45,000+. The highest leverage move isn't using Openclaw. It's learning to build your own single-purpose agents that outperform it 5-10x on your specific use case. I call these sniper agents. Full breakdown ↓

@_alphashark_ More tokens, more LLMs!!!!

@AgenticLab1 That breakdown of Openclaws architecture is fascinating! The integration of heartbeats and cron jobs adds a whole new layer of functionality. How do you envision these building blocks evolving in future AGI applications?

Full video: youtu.be/Bo4Shk2FCvk

@jrswab psst: parallel exploration subagents for mapping and tracing

cognitive debt is a skill issue. I too have this skill issue but am actively finding ways to increase my understanding as the speed of code generation increases.

@banteg Break glass in case of emergency: Type /model sonnet[1m] and type /compact then type /model opus Fixed!

what am i supposed to do with this? seems an unrecoverable state in claude.

@melvynx terrible advice speedrun

PRO TIP: Unlock the biggest update for your Claude Code now: 1. Go to .claude/settings.json and add this: "env": { "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-sonnet-4-6-1m", "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-6-1m" } 2. Use /model to switch to Sonnet Enjoy a 1M token context window and build an entire SaaS without auto-compaction...

@steipete Can't you just embed and discard if >similarity threshold? And ralph loops for deep reviews?

PRs on OpenClaw are growing at an *impossible* rate. Worked all day yesterday and got like 600 commits in. It was 2700; now it's over 3100. I need AI that scans every PR and Issue and de-dupes. It should also detect which PR is the based based on various signals (so really also a deep review is needed) Ideally it should also have a vision document to mark/reject PRs that stray too far. This can't be fully automated, but even assisting would help. The closes I found is an obscure oss project. How's no startup working on this?

Comment 'agents' and I'll send you my 16 page advanced claude code guide to help you get started on building these systems.

Code is worthless now. Here's what actually matters. Many people have the framing that code is inherently valuable. In the post coding agent world, this is no longer true. Specs are the true source of value, and the systems you build to turn those specs into working software are what separate people who vibe code from people who engineer at 50X speed. This means code repositories have become tuneable and portable. If you have a bug-ridden mess, you can scrap it, keep your specs, and have the agents rebuild it. It's crazy to live in a world where simple markdown files can be more valuable than gold. Think of it like a pop-up tent. Your spec and implementation plan are the tent in the bag. Your coding agent unfolds it, it pops into an app. If something goes wrong, you can just fold it back down, adjust the foundations, pop it up again. The spec is the thing you actually tune, not the code. But here's the part that most people skip out on: the mechanism that pops that tent out matters just as much as the tent itself. Get the pop-out wrong and you get a mangled tent, and if you don't have a mechanism at all, you just have a pile of metal sticks and some cloth. That mechanism is how you work with your coding agent. It's your slash commands, your context engineering, your orchestration patterns; how you feed specs to the model, how you manage subagents, how you structure your AGENTS.md/CLAUDE.md so the agent gets the information it needs. This is the craft now. This is what separates vibe coders from builders operating at 50X speed. It is a huge focus of mine to learn development in this style early, because all of this will become more and more true as models get faster, cheaper, and better. This changes the entire hierarchy of what you should be developing: 1. Read specs to understand intent. Read tests to understand behavior. Read code only when debugging gaps between the two. 2. Give yourself permission to scrap buggy code. Keep your specs, tune them, and rebuild. 3. Invest as much time learning how to work with your agent as you do writing specs. Your slash commands, context management, and orchestration patterns are the mechanism that makes everything else work. 4. Learn to build your own systems. There is no one-size-fits-all. Learn the foundations, then build what fits your workflow. Comment 'agents' and I'll send you my 16 page advanced claude code guide to help you get started on building these systems.

@yacinelearning I went ahead and debunk the concept for ya. youtube.com/watch?v=MYt9zl…

it has started…

@matt_elevenlabs 11B

We just raised $500M at an 11B valuation 🎉 To celebrate, we’re giving away 1,000 free credits so you can test our platform. For the next 6 hours, comment “11B” below and we’ll DM you the credits (must follow) 👇

New video! This should help you get up to speed with how to use subagents effectively. youtu.be/clCQCSZNSgA

@srinitude @GeoffreyHuntley Heck yeah man that is freaking sick!!

@AgenticLab1 @GeoffreyHuntley Celebrating your Geoff endorsement while I’m at NAMM rn :)

this is one of the best explainer videos for ralph that i have seen. declaring the official explanation m.youtube.com/watch?v=I7azCA…

@srinitude @GeoffreyHuntley Haha this is my guy!!

@AgenticLab1 @GeoffreyHuntley Let’s go Roman!