AI News Roundup: October Eighteen - Agents Hit the Wall

Overview: The agent economy ran into its first serious wall today. Two of America's largest brokerages confirmed they block outside AI agents from touching customer accounts, Amazon and Yelp kept their own doors shut, and Meta quietly deleted the contract language that would have let it scrape the web for training data. At the same time the money moved toward people: Anthropic carved 5% of its equity into a public benefit fund, Disney bought Character.AI's engineers out of bankruptcy, and Google published research showing the harness around a model matters more than the model itself.

JPMorgan and Schwab Are Locking AI Agents Out of Accounts

Schwab CEO Rick Wurster said his firm now receives a weekly report of blocked agent attempts on client accounts, describing them in the "tens and twenties" and naming Meta's Muse as one example. JPMorgan's Jamie Dimon confirmed his bank is blocking outside agents too. Schwab is redirecting users to its own Charley assistant and standing up Model Context Protocol servers that give external models read-only access while prohibiting transfers.

The pattern now spans sectors. Amazon has kept Meta Muse out since September 20, Yelp only admits agents enrolled in its paid data-licensing program, and Delta has signed no third-party agent deals. This is private ordering: regulated companies writing their own agent-liability rules while federal agent legislation stalls in committee. For builders, the practical lesson is that a browser-automation agent is a party nobody agreed to negotiate with.

Meta Pulls AI-Training Scraping Clause From Business Terms

Meta removed a clause from its commercial terms that legal experts said allowed it to scrape any site tied to a business account and use that content to train its models. The language had drawn fire from publishers and researchers who read it as a licensing-fee workaround. According to The Verge, the removal came after weeks of reporting on the clause.

The rollback matters less for what it says about Meta's intentions than for what it reveals about enforcement. When platform terms can change twice in a month, the only machine-readable opt-out that survives is the one you control: robots.txt, rate limits, and authentication on your own endpoints.

Disney Hires Most of Character.AI's Workforce After Bankruptcy

Disney agreed to hire at least 60 Character.AI employees for roughly $60 million a year, days after the companion-chat startup filed for Chapter 11. Disney is spending a fraction of that on the company's intellectual property, roughly $20 million, which tells you exactly what it was buying. In bankrupt AI application companies, the durable asset is the team that knows how to make character models feel alive, not the product or the user base.

Google's Harness² Proves the Scaffold Beats the Model

Google Research published Harness², a recursive agent harness that lifted Z.ai's open-weight GLM-5 by 11.5 points on WorkBuddy-Bench without touching a single model weight. The trick is recursion: the harness rewrites its own prompts and tool-use policies between attempts, encoding what worked into the scaffold instead of hoping the next model generation fixes it.

The finding lands the same week Anthropic disclosed that AI-agent work hours inside the company outnumber employee hours six to one. Both point the same direction. The upgrade surface that actually ships is the harness: prompts, skills, guards and tool routing, all versioned as code.

Anthropic Commits 5% of Its Shares to a Public Benefit Fund

Anthropic said it will place 5% of its shares into a wholly-owned entity called the Anthropic Public Benefit Fund, with the option to add another 2% for every 50% rise in company value above a baseline. Combined with co-founder pledges of 80% of their equity and a staff donation match, roughly a fifth of the cap table is now earmarked for public good.

The company floated large donations to nonprofits and direct distributions to specific populations, and tied the fund to its $200 million economic-futures program on labor-market disruption. Governance details are promised "in the coming months," which is the part that decides whether this is philanthropy or a press release. The timing is not subtle: CEO Dario Amodei has called public distrust of AI "fundamentally a crisis of trust," and Anthropic is heading toward an IPO.

What to Watch Tomorrow

• OpenAI's humanoid timeline: Altman confirmed on the Sources podcast that OpenAI will build a humanoid and expects a demo in 2027, with 27 robotics roles open including four in actuator design, per Engadget.
• OpenAI's unannounced "aurora" models: three undocumented model IDs surfaced in the API today, with one returning capacity errors rather than 404, suggesting a backend already wired up.
• Samsung hands HBM base dies to TSMC: Samsung will reportedly use TSMC's 3nm process for Nvidia custom HBM base dies, the first time it has outsourced that step.

Builder Impact

• Design for refusal. The default state of the open web and regulated finance is now "no agents." Build sanctioned interfaces first, treat browser automation as a documented fallback, and scope every action so a counterparty can accept it.
• Own your opt-out signals. Platform terms are a moving target. robots.txt, per-endpoint auth and rate limits are the only controls that do not get rewritten by someone else's legal team.
• Invest in the harness, not the wait. An 11.5-point gain from scaffold work is cheaper and faster than a model migration, and it compounds.
• Talent is the acquirable asset. Character.AI's bankruptcy made the point: teams transfer, products do not.
• Watch the governance, not the pledge. A public benefit fund with undisclosed decision rights is a promise. The structure, published in the coming months, will be the actual news.