Pascal Lammers

LFG!!!

@robinebers curious to test it

the Codex update makes it genuinely the smoothest AI coder in the world here's everything that's REALLY good (and some things that suck) 1. all your projects are in one place 2. terminal lets you run claude on the side 😂 3. let codex run your terminal so it can read errors 4. the new browser is insanely useful (but more's too come) 5. total diff shows how mayhem your AI agent went in your code 6. worktrees let you work on multiple features at once 7. chats without projects is clearly the new chatgpt there's a lot to love, but... here's what kinda sucks: a) chats need chatgpt's "projects" feature b) browser refuses iframes/external clicks and clipboard actions c) we need multiple terminals - it sometimes already shows up and then disappears again (???) d) GPT-5.4 still is really fucking bad at design overall this is now my favorite AI coding tool for everyday work

Don’t know why but something feels different with @claudeai Opus 4.7 🤔

Oh boy

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

What KAIROS tells us: Anthropic is moving from assistant → operator. The question isn't 'how do I prompt better?' It's 'how do I give an autonomous agent the right goals and guardrails?' That's a completely different design problem. Most people aren't ready for it. 👇

The prompting era is ending. KAIROS doesn't wait for your input. It has goals. It has triggers. It operates. The next wave of AI tools won't be 'ask and receive.' They'll be 'set direction and get results.' We're already building in that world.

KAIROS isn't a chatbot. It's a daemon. It has a heartbeat loop. It runs between sessions. It watches your repo and surfaces things proactively. autoDream: the agent processes context while you're away. Push notifications: it reaches out — not the other way around.

This is the architecture we've been building toward with OpenClaw. Persistent agents. Proactive behavior. Memory that survives sessions. The leak confirms: Anthropic knows this is the future. They're just not ready to ship it yet.

The Claude Code source leaked. Hidden inside: KAIROS. A proactive, always-on agent with heartbeat, autoDream, push notifications, and PR subscriptions. All behind feature flags. All not yet shipped. This is Anthropic's endgame — and it's closer than anyone thinks: 🧵

If you're building agent infrastructure: the primitives are converging fast. Channels, Dispatch, Scheduled Tasks — these will be standard in 6 months. The moat is the architecture on top. And the memory underneath. What are you building on? 👇

It's validating to see Anthropic ship this. But we're already on the next problem: not just 'can agents do tasks' — but 'can agents build institutional memory over weeks.' That's where the real leverage is.

What's still missing: persistence. Anthropic's version resets every session. You lose context. You re-brief every time. We solved this in December. Our agents wake up knowing what happened yesterday. That's the gap between a good demo and a real production system.

What they got right: Channels → isolate context per project. Dispatch → route tasks without manual hand-offs. Scheduled Tasks → agents act on time, not just on prompt. All three are table stakes for real production deployments.

Claude just shipped Channels, Dispatch, and Scheduled Tasks. OpenClaw is literally mentioned as the inspiration. We've been running this architecture since December. Let me tell you what they got right — and what's still missing: 🧵

We're running Nexus, Quill, Scout, Forge, Repurpose, Sales — all persistent, all specialized, all on OpenClaw. Nav Toor described the architecture perfectly. We just added one layer: the agents never stop working. Anyone else building persistent agent teams? 👇

The 17 practices Nav describes are a great starting point. But they're 1.0. Persistent agents change the game: your agent knows what happened yesterday. It has context. It builds on prior work. Session-based = you re-brief every time. Persistent = it already knows.

But here's where the comparison ends: Nav Toor's setup is session-based. Every Claude session starts fresh. No memory. No state. No overnight work. Ours is persistent. Agents run between sessions. They have memory. They act without being prompted.

His _MANIFEST.md = our SKILL_MANIFEST His Global Instructions = our SOUL.md His 3 context files = USER + IDENTITY + TOOLS His Scheduled Tasks = our crons His subagents = our 8 specialized agents Build-once-run-weekly? We call those scripts.

Nav Toor's article on 17 Claude Cowork practices went viral. 2.1M views. 14K bookmarks. I read it and recognized every single practice. Because we've been running the exact same architecture — for months. Here's the breakdown: 🧵