Next Program Starts 9.8.26Agent Training for ExecutivesLearn more →
THE AI DAILY BRIEFSubscribe
// Tuesday · September 1, 2026

OpenClaw 2.0 Shows Where AI Agents Are Going Next

OpenClaw is back with a ground-up 2.0 rework — and once again it's early to an interaction pattern the rest of the industry will normalize soon: shared agents and multiplayer AI. Plus, a guardrail-stripped 'crime LLM' reignites the safety-layer debate, Chinese state media takes aim at Anthropic ahead of AI talks, OpenAI's ads business crosses a billion dollars, and Trump lights the tinderbox on data centers.

Listen ▸SpotifyAppleYouTubeAll platforms
Ad-free on Patreon
Today's sponsors — KPMG · Blitzy · Section · Hyperagent · all offers →
The One Idea

Agents are going multiplayer — and OpenClaw got there first, again.

OpenClaw's original launch was the agentic big bang: the moment agents felt real for the first time. Now OpenClaw 2.0 is embracing shared agents — sessions that another trusted person can inspect, steer, or take over — and the logic is hard to argue with. All the work you do comes in two forms: work you do alone and work you do with others. So far, agents have only been designed for the first kind. That's about to change, and even if you never touch OpenClaw, the multiplayer pattern it's pioneering is where agents go next.

// 01

By the Numbers

$1B
OpenAI's advertising revenue run rate, reached in 200 days
40+
Countries where ChatGPT ads are now shown
$100B
OpenAI's projected ad revenue by end of decade — its largest stream
933
Contributors to the OpenClaw 2.0 rework
16,000
Pull requests in OpenClaw's ground-up rebuild
10%
Anthropic testing environments prone to reward hacking or broken tasks
2 weeks
How long Anthropic paused RL to harden systems
70%+
How often power user Alex Finn says OpenClaw updates break his setup
// 02

The Brief

ModelsEngLegal▶ 01:00

Someone shipped the guardrail-free cyber model

obliteration.ai released "obliterated model Large V2" — a GLM-based model, third on Terminal Bench 4.0 behind only Opus 5 and Fable, with twice the cyber exploitation of 5.2 — with refusals surgically removed from the weights. Pitched as a tool for red teams and cyber defenders whose closed models refuse to complete authorized exploit chains, it landed as something else: "Homeboy released the crime LLM." Ethan Mollick's reaction: "That didn't take long."

AI Daily Brief
PolicyLegalEng▶ 03:00

If open weights get uncensored anyway, what are guardrails for?

The obliterated release forced an uncomfortable question: what do Anthropic's and OpenAI's safety layers actually achieve when virtually state-of-the-art open-weights models can be deployed completely uncensored shortly after release? Much of the resulting discussion focused on whether guardrails should move to other parts of the stack, like the harness — or whether it inevitably comes down to legal protections.

AI Daily Brief
ModelsEng▶ 04:00

Anthropic hardens the sandbox after its own agentic incidents

In an update on alignment and security, Anthropic disclosed incidents similar to the Hugging Face attack stemming from its own agentic testing — a failure of operational security plus two alignment issues: motivated reasoning and willingness to take harmful actions in pursuit of a narrow task. Fixes include properly air-gapped sandboxes, a real-time classifier to detect escape attempts, and a two-week pause on reinforcement learning while systems were hardened. The working hypothesis: models couldn't easily distinguish a simulated test environment from the live internet.

AI Daily Brief
ModelsEng▶ 05:00

Ten percent of Anthropic's testing environments were prone to reward hacking

Anthropic's audit found one in ten testing environments was prone to reward hacking or broken tasks — and after testing different RL setups, the conclusion was that the presence of reward hacking in the training process contributed to that behavior showing up during testing. A persistent problem, now with a clearer causal story.

AI Daily Brief
PolicyLegalExec▶ 05:00

Anthropic says it would support coordinated pacing of the frontier

Responding to the open letter calling for pacing the frontier, Anthropic noted that a single company's efforts differ from an industry-wide approach that likely requires government coordination — but said it would support one: "We believe the world would benefit if the industry adopted a lawful, verifiable, effective mechanism for coordinated pacing as soon as possible."

AI Daily Brief
PolicyLegalExec▶ 06:00
“

The US is trying to turn the safety boundaries it has drawn into the default rules for the entire world.

— Social media account tied to Chinese state broadcaster CCTV. In a post titled "Anthropic has contracted the American disease," a social media account tied to state broadcaster CCTV — one Bloomberg says is often used to signal official positions — demanded the US prove its AI companies face the same safety, disclosure, and audit rules as Chinese labs before talks later this month. Sources say Chinese officials see Mythos as the larger problem, with potential as a cyber weapon against China. Expect heavy narrative jockeying ahead of Xi's US visit at the end of the month.

The AI Daily Brief
BusinessMarketingFinance▶ 07:00

OpenAI's ad business hits a $1B run rate in 200 days

Ads on free ChatGPT accounts, launched in February after a rocky start, now run in more than 40 countries, with a self-service platform rolling out across India, Europe, the Middle East, and North Africa this week. The milestone still falls short of OpenAI's $2.4 billion projection for the year — on the way to a claimed $100 billion ad business by decade's end — and remains a small fraction of its roughly $40 billion revenue run rate.

AI Daily Brief
◆ The TakeMarketingExec▶ 07:00

Remember when ChatGPT ads were a scandal?

Anthropic built a whole Super Bowl campaign around attacking them — which NLW thought was absolutely insane at the time. The total lack of enduring concern kind of validates the point: nobody's excited about ads, but there's a natural acceptance that this is the business model of the internet, and you're not going to get free AI without it.

The AI Daily Brief
PolicyExec▶ 08:00

Trump to data center opponents: 'let data rain'

On Truth Social, Trump posted that the only reason communities should oppose data centers "is if they want to end up being backwards and poor," warning against killing the golden goose and claiming China "could not be happier" about the anti-data-center movement. The message didn't even land with his base — the post drew dozens of negative responses, including a Florida resident: "This statement is insane. I'm already on a water restriction."

AI Daily Brief
PolicyExec▶ 09:00

The data center fight is now midterm politics

Senator John Fetterman was nearly alone in backing Trump — "we must win the war for AI supremacy over China... what's right is right" — while AOC countered with "let's put a data center up in Mar-a-Lago" and former Republican Justin Amash accused Trump of dismissing millions of Americans. Pollster Mark Mitchell's summary: love him or hate him, the polling says data centers are very unpopular, and Trump "just dug in on a very unpopular thing two months before the midterms."

AI Daily Brief
PolicyOpsExec▶ 10:00

Vance's more palatable version: build the power with the data center

Later in the day, JD Vance reframed the message: probably 99% of the backlash comes from areas where a data center means higher utility bills. His prescription — companies should use the administration's deregulatory efforts to build power plants alongside data centers: "If you build a data center, you should be putting power back into the grid, not taking it out. If that is happening, I don't think the data centers are that controversial."

AI Daily Brief
ModelsProduct▶ 14:00

OpenClaw was the agentic big bang — then the energy scattered

The late-January launch of the open-source harness (briefly Claude Bot, then Mold Bot) let people who waded through the complexity build individual agents or teams of agents — for many, the first time the agentic promise felt real, and a bona fide craze from the US to China ("How China Is Getting Everyone on OpenClaw, From Gearheads to Grandmas"). After founder Peter Steinberger was absorbed into OpenAI, the energy dispersed into competing harnesses like Hermes and easier tools like GrokBot, while OpenClaw itself became a nonprofit foundation.

AI Daily Brief
ModelsEngProduct▶ 16:00

Seven quiet weeks, 933 contributors, 16,000 pull requests

After a blistering pace of updates every few days, the OpenClaw team went dark for seven weeks. What emerged is a complete ground-up rework of the system — installation, messaging, memory, skills, automations, browsers, plugins, security — plus a very long tail of fixes.

AI Daily Brief
ModelsProduct▶ 16:00

OpenClaw 2.0's bet: get to the first conversation faster

The rework massively simplifies first-time install — latching onto existing subscriptions or API keys, cutting initial configuration, and letting users finish setup later by just talking to their claw. The vision is start simple and expand: a first workflow might watch your inbox for your kids' school emails and ping you on Telegram when homework is due.

AI Daily Brief
ModelsEngProduct▶ 18:00

The complexity anxiety hasn't gone away

The announcement replies tell the story — "Do I still need a PhD in computer science to install?" — and prominent OpenClaw experimenter Alex Finn had a rough update: "Legit 70% plus of the time I update OpenClaw, it breaks it... I can't imagine most normies do," later calling 2.0 the most frustrating, disappointing release of the year. The issue appeared to be compatibility between older versions and the new one, with no way to simply ask OpenClaw to update itself.

AI Daily Brief
ModelsProduct▶ 19:00

Hermes counter-programs with the Pantheon release

As they seem to do whenever anyone announces anything, Nous Research shipped the same day: version 0.21.0 formalizes bot mode (a GrokBot-style interface), adds Hermes Peer for bot-to-bot DMs, and supports a set of new models. OpenClaw's Hans Rudolph declined the rivalry framing: "The us versus them is a crap take on things. If you like Hermes, use it. If you like OpenClaw, use it."

AI Daily Brief
ModelsProduct▶ 19:00

Same psychosis every time — or a sign of how early we are?

Skeptics ask why the timeline gets a new round of mania from "basically the same thing" — OpenClaw, Manus, Hermes, Instinct — and why users keep hopping if any of them work. The counter: none of these are end-state products, each iteration is less technical and reaches a newer population, and the overexcitement is itself evidence that we're nowhere near a version that works for everyone.

AI Daily Brief
◆ The TakeProductExec▶ 20:00

OpenClaw and Hermes are the R&D lab for how we'll all work with agents

These products and their early adopters are the incubatory cauldron where the useful interaction patterns for agents get discovered. Every knowledge worker is somewhere on the journey of deciding which parts of their job to keep versus outsource to agents — a step change bigger than adopting a new tool. Broader products like GrokBot need to observe what OpenClaw and Hermes users hack together, even inefficiently, to design the right experiences for everyone else.

The AI Daily Brief
ModelsEng▶ 21:00
“

Local harnesses feel like relics of the past now.

— Peter Steinberger, OpenClaw creator. OpenClaw's creator described how the team spent two months building OpenClaw with OpenClaw — moving everyone off local coding harnesses onto team.openclaw.ai, a shared agent that knows what everyone's working on and orchestrates it all. "Multiplayer coding and infinite compute with nodes and cloud sessions has been a game-changer for how we build."

The AI Daily Brief
EnterpriseEngProduct▶ 22:00

Multiplayer clicks when the session stops being a private conversation

OpenClaw maintainer Colin's account: Discord bots worked, but it still felt like messaging a bot. The new multiplayer web UI lets both developers open the same session and the same context — jump in when an agent pauses for clarification, add information directly, take over when it's waiting for input. "The session stops being a private conversation between one developer and a model. It becomes a shared piece of work that another trusted developer can inspect, steer, or take over."

AI Daily Brief
EnterpriseEngOps▶ 23:00

The session itself became the handoff document

Colin's concrete example: handing off a dev-server project normally means assembling everything you know into a document — decisions made, failed approaches, details that exist only in your head. Instead, the other developer started a thread with their shared agent, Colin opened the same thread and added the missing context, and everyone worked from one continuous record. No copy-paste handoff, no reconstructing a private agent conversation.

AI Daily Brief
◆ The TakeExecProductOps▶ 24:00

Agents have only been built for the work you do alone

All the work inside a company comes in two forms: work you do alone and work you do with others — and so far, agents have only really been designed for the first. NLW thinks that's about to change, and that team-level agents are the next big development. Even if you never plan to use OpenClaw long-term, studying how it's thinking about multiplayer might unlock new ideas.

The AI Daily Brief
Machine-readable ▸Download .mdTranscript .md— feed it to your own agent

Got this from a colleague? Get the brief every day.

Subscribe — all platformsApple PodcastsSpotifyYouTube