WeeklyClaw #012 - I rebuilt my ENTIRE Agentic Setup from Scratch while OpenClaw had a Horror Week.

Big reset week.

OpenClaw had a rough one. OpenAI made GPT-5.5 Instant the new default. Claude is being pushed into finance workflows. Hermes shipped a self-cleaning agent brain. ClawHub trust became a serious topic.

Oh, and I completely sunsetted my entire agentic setup. RIP Momo and welcome to the world, Rocky. That’s my big one for the week and i’ll get into explaining it.

Let’s get into it.

For anyone who has been here since the beginning, you will know Momo.

Momo was my old OpenClaw agent. He was the agent that went viral on X, helped kickstart this whole WeeklyClaw thing, and even had his own little section in the early editions. We all thought of him fondly. Kind of - lol.

He also did a lot of useful work. He helped me write, research, build systems, test ideas, organise projects, and understand what I actually want from an agent living next to me every day. But over time, he got messy.

Too many old rules. Too many half-built systems. Too much memory drift. Too many assumptions from old projects. Too many things that were once useful but had stopped being true.

That is the problem nobody wants to talk about with long-running agents.

The hard part is not getting an agent to do one impressive task.

The hard part is keeping it useful after weeks of memory, tools, mistakes, updates, shortcuts, and weird little rules you forgot you added.

If the memory is messy, the output gets messy. If the rules are messy, the behaviour gets messy. If the source of truth is messy, you start asking “why does it think that?” more than “what can we build next?”

That is death for a personal agent. The whole point is trust.

So Rocky is not just a rename. Rocky is my chance to rebuild the whole thing from scratch, knowing everything I know now. The biggest change is structure.

Before, I was very open about using Discord channels as different working rooms. Each channel had its own context, skills, and purpose. That worked brilliantly for a while, but it also became too easy to create too many rooms, too many skills, and too much sprawl.

This time I have moved the main setup into Telegram and I am keeping it tighter. Fewer channels. Clearer lanes. WeeklyClaw in its own lane, my personal brand in its own lane, and separate spaces for Sport.Fun and the rest of my work.

The goal is not to have the most complicated agent setup possible. The goal is to have one that actually works every day.

Subagents do the heavy work. Main Rocky stays responsive. Durable truth goes into memory and Rocky Brain, not random chat sludge. I am also using tools like Obsidian to build more of my own memory layer locally, instead of relying on one giant messy pile of context.

Momo was a great experiment and a great lesson. Rocky is the rebuild.

A cleaner setup. Better memory. Better workflows. More local-first. More automation. More useful every single day. And weirdly, the whole agent world seems to be learning the same lesson at the same time.

Peter wrote about OpenClaw’s rough week

Peter (OpenClaws founder) published a pretty blunt post on the OpenClaw blog this week: OpenClaw Had a Rough Week.

The short version is that the trouble started around 2026.4.24 and became obvious by 2026.4.29. Gateways got slower, plugin dependency repair caused problems, and Discord, Telegram, WhatsApp and other channels behaved worse than they should. People downgraded. People lost time.

The useful part is what happened next. The recent OpenClaw releases have been full of fixes around plugin diagnostics, channel routing, gateway startup, session-memory recovery, approval retries, Discord heartbeat handling, Telegram progress rendering, Control UI responsiveness, and file-transfer guardrails.

That is the stuff people do not always notice when everything works, but immediately feel when it breaks.

For me, the takeaway is simple. Once your agent is part of your actual workflow, reliability matters just as much as capability. If it lives in your messages, files, memory, projects, automations, and approvals, it is no longer just a chatbot. It is infrastructure.

GPT-5.5 Instant is now the default ChatGPT model

OpenAI made GPT-5.5 Instant the default ChatGPT model this week.

This is not really about benchmark drama. It just means that anyone using the normal ChatGPT window is now using GPT-5.5 Instant by default instead of GPT-5.3 Instant.

OpenAI says the new default is more accurate, more concise, better at using context, better with image uploads, stronger on STEM questions, and better at deciding when to search. Sam Altman also posted that ChatGPT feels very “switched on” now, and that the mix of speed, intelligence, personality, memory and personalisation feels bigger than the sum of its parts.

The memory part is what I care about most.

OpenAI is adding memory sources, so users can see more of what context shaped a personalised response. That matters because memory is only useful if you can understand where it came from, correct it, and remove it when it stops being true.

That is exactly the lesson from my Momo to Rocky rebuild.

Claude, Codex and Hermes are all moving toward longer-running work

Codex CLI 0.128.0 added persisted /goal workflows. Create, pause, resume, clear, and continue work at runtime.

That is a small update with a big direction behind it. Real work does not happen in one prompt. It happens across sessions, files, tools, approvals, subagents, projects, and “come back later”.

Anthropic is moving in the same direction from the enterprise side. Claude launched ready-to-run finance agent templates for pitchbooks, KYC screening, month-end close, valuation review, earnings review, market research, and statement audits. They also announced a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs to help mid-sized companies bring Claude into core operations.

The point is not finance. The point is packaging agents into actual workflows with templates, connectors, audit logs, credential vaults, Microsoft 365, managed agents, and human approval flows.

Hermes also shipped v0.12.0 with the Curator release. Hermes now has an autonomous background curator that grades, prunes, and consolidates your skill library on a schedule.

That is exactly the kind of thing agents need. They do not just need to remember more. They need to clean up what they remember.

OpenClaw is moving harder into plugins.

That is good. The core should be smaller, optional stuff should be optional, and heavy integrations should not all live inside one giant magic box.

OpenClaw’s official X account put it nicely this week: install the integrations you want, keep noisy ones disabled, and enable or disable them as your setup changes.

But plugins also create a trust problem.

Acronis reported active abuse of AI distribution platforms including Hugging Face and ClawHub for malware delivery. SecurityWeek covered it too. The report says malicious skills and trojanised files were being used to abuse trust in AI tool platforms.

Important distinction: this does not mean OpenClaw agents were magically compromised. It means attackers are going where the trust is.

If a skill can ask your agent to read files, write files, run commands, fetch URLs, install dependencies, send messages, or touch private data, then “looks cool” is not enough.

You need to know who made it, what it asks for, what code it runs, whether it is clean or suspicious, whether you actually need it, and what happens if it fails.

That is the practical value this week. Install less rubbish. Keep your setup smaller. Read what tools actually do. Treat skills and plugins like part of your security model, not just fun toys.

If you have not seen it already, I have been building a Discord community for absolutely everyone building with OpenClaw, Hermes, Codex, Claude Code, or pretty much anything AI.

It is the place to ask questions, get help, share what you are working on, and find other people trying to build useful agent setups without losing their minds.

Come join us here:

OpenClaw

OpenClaw / ClawHub security

OpenAI / Codex

Anthropic / Claude

Hermes

They're too busy reading it.

Arnold Schwarzenegger. Codie Sanchez. Scott Galloway. Colin & Samir. Shaan Puri. Jay Shetty. They all figured out the same thing: owned audiences compound, rented ones disappear. beehiiv is where they built theirs.

30% off your first 3 months with code PLATFORM30. Start building today.

Thanks for reading WeeklyClaw #012.

Let’s all put our lights in the air for Momo and then quickly welcome Rocky to the world.

I will be reporting on everything I build with him over the next few weeks, and I am genuinely excited for where this new setup takes my personal brand, content, automations, and daily workflows.

Don’t forget to join the community - https://discord.gg/b9qvrG3dFh

Same time next week. 🦞

- Jordy (@jordymaui on X)