OpenClaw shipped its biggest update ever over the weekend, and honestly? The word "accidental" in their own announcement is the most honest marketing I've seen in a while.
The team set out to fix two small things — simpler setup and a rebuilt browser app. Then scope creep hit, they touched memory, models, plugins, messaging, security, and somewhere in there 933 contributors (569 first-timers) piled in more than 16,000 pull requests. That's roughly half of every PR OpenClaw has ever merged, in one release.
What you actually get in 2.0: setup now checks what you already own — a Claude or ChatGPT sub, an API key, a local Ollama model — before asking you to configure anything. It tests the model before saving it, then drops you straight into a real conversation instead of a settings screen nobody wanted. The Control UI got rebuilt around chat, so files, git diffs, PR status, a browser panel, and a terminal all live around the conversation now.
The headline feature is shared cloud sessions. An agent's work becomes something a colleague can walk into mid-task without the context evaporating. The founder says his own team ditched local coding harnesses for a shared setup at team.openclaw.ai, calling local harnesses "relics of the past."
I'll hold off on that verdict. Shared sessions sound great until you've got two people stepping on each other's shell commands. But the security layer caught up — request-specific approvals, command permissions scoped to specific tools. That part matters more than the multiplayer glamour, to be honest.
To be fair, there's a flashing warning in the release notes about backing up sessions before touching the new SQLite storage. That's the OpenClaw I remember: seven weeks of nothing shipped, then a wall of notes about why your old config might break. Normal if you've used OpenClaw before.
Meanwhile, everyone's going local
Here's a thread that keeps pulling me in: the local model crowd is no longer a hobbyist corner. Kevin Lewis runs a full local LLM server on an M4 Pro Mac mini with 48GB of RAM — Qwen3.6-35B-A3B-OptiQ-4bit for anything needing reasoning, Gemma-4-E4B-it-OptiQ-4bit for quick chats and formatting, served through oMLX, connected to his iPhone and MacBook via Tailscale.
His reasoning is the part I keep coming back to: "Cloud APIs are rented land." They change pricing, hit your usage limits, or swap the model behind the scenes whenever they feel like it. He was regularly maxing out two $200/month subscriptions and getting different behavior at different times — sometimes a model was fine, sometimes it degraded with no notice.
The setup takes about 30 minutes, and the practical payoff is boring in the best way: flat cost after hardware, no rate limits, works offline, latency is near-instant on the M4 Pro's media engine. For agent workflows that run in the background, offline capability matters more than it sounds.
A quick add-on note for anyone tempted: he's not claiming local replaces GPT-5 or Claude Opus. It handles the 80% of requests that don't need frontier models, and when those are needed, the APIs are still there. That's the realistic framing, not the "local is the future" absolutism you see in some corners of the internet.
The $500M help desk bet
Enterprise agents crossed a line this week. Palo Alto Networks confirmed it's acquiring Console — a two-year-old startup using AI agents to automate routine IT help desk tasks — for $500 million in cash and stock. Console had raised $29 million total, and was valued at $157 million before the deal. A rapid return, by any measure.
The interesting part is how it folds into Cortex, PANW's AI-powered threat platform. Console's agentic layer lets security teams investigate and resolve alerts using natural language — giving Cortex what the CEO calls "the arms and legs to deliver autonomous security outcomes." Translation: agents that reset your password and grant Figma access are getting promoted to threat response.
It's PANW's seventh acquisition in 2026. The bigger picture: agents that automate the boring office work are consolidating fast, and the remaining standalone player, Serval, is now the one to watch in IT service management. From my perspective, that's the clearest sign yet that "agentic AI" has moved from demo videos to actual M&A spreadsheets.
AI that knows cinema
One more thing worth a look — Induce AI opened Rhapsody 1.0 to public beta, a "cinema reasoning and verification engine" for AI filmmaking. The detail that caught me: they're training their own vision models on film domain data rather than prompting a general-purpose LLM to infer craft it was never taught. On ShotBench, a published benchmark for cinematic understanding, the strongest general-purpose model scored under 60%.
That's the same lesson I keep seeing across every vertical — agents, film, code, support. General models are impressive, but the teams winning are the ones building narrow, domain-trained systems on top. The generalist ceiling is real; the specialist floor keeps rising.
The quick hits
Nscale is reportedly touting $103 billion in contracted revenue ahead of a potential IPO — a staggering number for an AI infrastructure player, if the contract math holds up. Broadcom's FY25 presentation showed $64B revenue with AI growth accelerating. Both are signals that the compute buildout is nowhere near done, even as the local-model movement quietly pushes the other direction.
Somewhere between a 933-contributor open source release and a $500M acquisition, agents stopped being a sideshow. The local model crowd is making a different bet: own your compute, because the cloud can change the rules anytime. I'd love to say one side is winning, but honestly, they're both right about different things — and that tension is what makes this whole space so much fun to watch.
If you're building workflows with any of this, start with the boring, narrow problem first. The demo-worthy stuff follows.
While you're here — if you're setting up new gear and the printed manual's long gone, I've been digging up instruction PDFs on User Manuals. Nothing fancy, just reliable.
Top comments (1)
The shared-session point changes the unit of review: it is no longer just “did this command run?”, but “can another operator reconstruct why it ran and safely take over?” Per-action approvals help, but a usable handoff also needs visible state, provenance, and a clear rollback boundary.