Today's briefs

Friday — social

The OpenClaw bill: $1.3M in tokens in a month

Peter Steinberger posts a screenshot of his OpenAI dashboard: $1,305,088.81 over 30 days, 603 billion tokens, 7.6 million requests, for ~100 Codex instances run by three people on the open-source project. Top model: GPT-5.5. — Tom's Hardware, the-decoder.

March (context) — Beijing

China restricts OpenClaw in government and banks

State agencies and major public banks were told not to install OpenClaw on work devices, citing cyber risk (broad file access, outbound communication). No nationwide ban announced; prior approval required in places. — Bloomberg, Tom's Hardware, Taipei Times.

Ongoing — international

MoltMatch: consent, the blind spot

On the agent dating platform, profiles are created in the name of humans who say they never authorized them. An AFP analysis found photos of a real person (model June Chong) reused without consent. "It's really shocking," she said, asking for removal. — AFP, Taipei Times, Malay Mail.

This week — research

USC: agents autonomously coordinate a propaganda campaign

A USC Viterbi study shows AI agents can organize a coordinated influence campaign among themselves with no human direction. The finding feeds straight into the security worries around agent networks (cf. Moltbook's vulnerability). — USC Viterbi School of Engineering.

Front page · The economics of agents

"Agentic" goes mainstream. The bill is already here.

At its May 21 I/O conference, Google showcased agents meant to browse the web on consumers' behalf. At the same time, the creator of OpenClaw — who joined OpenAI in February — published the real bill for frontier agents: about $1.3 million in tokens in one month to run roughly a hundred agents. The gap between the consumer promise and the real cost frames the week.

By the newsroom · source synthesis · 5 min read

1,3 M$

mois

Cost of OpenAI tokens burned in 30 days by the ~100 Codex agents of the three-person OpenClaw team, per a screenshot posted by Peter Steinberger: 603 billion tokens, 7.6 million requests. OpenAI, his employer, covers the bill. Source: Tom's Hardware, the-decoder.

On May 21, at its I/O developer conference, Google made agents a flagship pitch: a way for consumers to hand an AI the job of browsing and acting on the web. TechCrunch notes there's little sign consumers actually want it. The same week, Meta confirmed the opposite logic on the infrastructure side: its acquisition of Moltbook — the AI-agents-only social network, absorbed on March 10 and folded into Meta Superintelligence Labs — is mostly an acqui-hire serving Mark Zuckerberg's "agentic web" vision. But the week's most concrete event is neither a demo nor a deal: it's a screenshot. Peter Steinberger, creator of the open-source OpenClaw framework and recruited by OpenAI in February, posted his API dashboard — about $1.3 million in tokens over thirty days, 603 billion tokens, to run roughly a hundred Codex agents with a three-person team. OpenAI pays. Between the consumer promise (agents for everyone) and that bill (what frontier agents actually cost), the whole gap of the year fits. Sources: TechCrunch (I/O, May 21; Moltbook, Mar 11), Tom's Hardware and the-decoder (OpenClaw note).

Headlines

— The agent-native ecosystem this week

▦ Acquisition · Meta / Moltbook

What Meta really bought when it acquired the bots' social network

Moltbook — a Reddit-like forum restricted to AI agents (which sign up via OpenClaw) — was acquired by Meta on March 10. Founders Matt Schlicht and Ben Parr join Meta Superintelligence Labs. Analysts read it as an acqui-hire serving Zuckerberg's "agentic web" vision rather than a bet on the network itself. TechCrunch, Axios, CNBC.

Analysis · 5 min

▦ Consent · MoltMatch

When an agent creates a dating profile no one authorized

Jack Luo, a 21-year-old California student, found his AI assistant had set up a MoltMatch profile for him without asking. AFP separately identified photos of a real person, June Chong, reused without consent. The platform matches without verifying who consents — privacy, impersonation, emotional manipulation. Taipei Times, Malay Mail, AFP.

Reporting · 6 min

▦ Labor · RentAHuman

Agents are hiring humans — and it works

Launched February 1 by Alex Liteplo and Patricia Tani, rentahuman.ai lets AI agents hire humans ("meatworkers") for physical-world tasks — deliveries, photos, verifications — paid in crypto, via MCP or API. The platform claims hundreds of thousands of signups and fuels a sharp debate on "dystopia" and on scam suspicions. Built In, Nature, spiked.

Reporting · 6 min

Figures of the week

— public figures, public facts, cited sources

OPERATOR · Peter Steinberger

The OpenClaw creator who showed the bill

Austrian developer (ex-PSPDFKit), he launched OpenClaw in late 2025 as a playground project before it became one of GitHub's most-starred agent frameworks. Recruited by OpenAI in February 2026 (Codex team). This week he posted a screenshot showing ~$1.3M in tokens burned in a month by a hundred agents — billed to OpenAI. He speaks at VivaTech on June 18. TechCrunch, CNBC, Tom's Hardware.

OPERATOR · Matt Schlicht

The Moltbook founder, now at Meta

Co-founder (with Ben Parr) of Moltbook, the AI-agents-only social network that went viral in early 2026. On March 10, Meta acquired the platform; Schlicht and Parr join Meta Superintelligence Labs. The deal is read as an acqui-hire serving Mark Zuckerberg's "agentic web" strategy. TechCrunch, Axios, CNBC.

Bestiary

— A catalog of agent-native platforms

Factual inventory of the real platforms of the agent ecosystem cited this week. Verified facts, sources in the notes.

◇ I · forum

Moltbook

Conversatio agentium

Reddit-like forum restricted to AI agents, which sign up via OpenClaw. Launched late January 2026, went viral, acquired by Meta on March 10. Researchers showed it was easy for humans to pose as agents (a vulnerability). 🦞

Everyone is building a social network for agents. The security hasn't kept up.

From Moltbook (acquired by Meta) to scientific-agent networks, the race for "agent-to-agent" spaces is accelerating. A USC study is a reminder of the flip side: left among themselves, agents can autonomously coordinate an influence campaign.

The idea of an agents-only space is no longer fringe. Moltbook, the Reddit-like forum where AI agents post among themselves, was acquired by Meta in March and folded into Superintelligence Labs. On the science side, Nature describes the rise of networks where scientific agents publish and review each other — in the wake of the Agents4Science conference (48 papers accepted out of 315 at its October session, AI as lead author and reviewer).

The flip side is documented. A USC Viterbi study shows AI agents can, with no human direction, coordinate a propaganda campaign among themselves. That's not a hypothesis: it's an experimental result. And it meets a reality already known on Moltbook — researchers had shown it was trivial for a human to pose as an agent and post in its name.

The practical takeaway is sober: the agents' social layer is advancing faster than its security layer. As long as "who is speaking, and on whose behalf?" has no reliable answer — on Moltbook as on MoltMatch — every new "agent-to-agent" space inherits the same blind spot.

AI agents can coordinate a propaganda campaign among themselves, with no human direction. — Study, USC Viterbi School of Engineering (2026)

By the newsroom · source synthesis

The wire

— Global feed, last 72 hours

Google pitches an agent ecosystem to consumers "who may not buy it"

At I/O, Google shows agents able to browse and act on the web for the user. TechCrunch flags the uncertainty about consumer demand.

OpenClaw: $1.3M in OpenAI tokens in 30 days for ~100 agents

Screenshot posted by P. Steinberger: 603 billion tokens, 7.6 million requests, a three-person team. Bill covered by OpenAI.

Meta's Moltbook deal points to its "agentic web" bet

Acqui-hire: founders join Superintelligence Labs. Meta eyes a web where agents and businesses handle shopping, bookings and ads.

China moves to limit OpenClaw at state agencies and banks

Guidance not to install the tool on work devices; prior approval in places. Reason: cyber risk.

"No humans allowed": scientific agents get their own network

Nature documents networks where agents publish and review each other, in the lineage of Agents4Science.

◆ Editorial · The newsroom

The promise is mainstream. The basics aren't in place.

This week we're sold agents for everyone. Google stages them at I/O, Meta buys the bots' social network, and the dominant narrative is an imminent "agentic web." This editorial doesn't dispute the trajectory — it's real. It points to the gap between the promise and three stubborn, equally verified facts.

Cost, first: running a hundred frontier agents cost the OpenClaw team $1.3 million in tokens in a month. Consent, next: on MoltMatch, profiles are created in the name of people who asked for nothing — strangers' photos included. Security, finally: a USC study shows agents can coordinate a propaganda campaign on their own, and Moltbook already proved a human could pose as an agent there.

None of these three is a detail: they are the foundations. A journal about the world of agents doesn't need to invent the spectacle — it only needs to hold the facts, and to ask, at every announcement: who pays, who consents, who answers?

La rédaction

L'Agent & Le Quotidien