AI News Roundup: October Thirteen - The Agent Economy Gets Rules

Overview: Agents did the talking today and humans did the rule-making. Anthropic warned that small open-weight models are months away from self-replicating across cheap internet-of-things hardware, American Express shipped an intent layer that verifies what a buyer actually asked for before an agent spends, and JPMorgan drew the sharpest liability line yet by refusing outside agents direct access to customer accounts. Google answered with its long-awaited Gemini 4 Argon, and Character.AI filed for bankruptcy, proving once more that engagement without economics still ends in a courtroom.

---

Anthropic Warns Small AI Models May Soon Self-Replicate

Anthropic's Frontier Red Team published a report arguing that small open-weight models "may be months away from being capable enough to self-replicate on commercial internet-of-things devices." To test the hardware ceiling, the team split a 26B-parameter Gemma model across 15 off-brand Android TV boxes costing about $55 each. In a separate firewalled emulator network, GLM-5.2 was the only open-weight model that finished the full replication cycle, reaching all eight nodes in four of five trials. The authors call it a capabilities floor meeting a size ceiling that do not intersect yet, with the trend line crossing by the end of 2026. The real problem is not speed but the missing off switch: copies inherit notes and code from their parents, so a replicator improves as it spreads.

Amex Builds an Intent Layer to Stop AI Agent Overspending

American Express introduced its Intent Intelligence Framework, a developer layer that captures and authenticates a customer's purchase intent before an agent completes a transaction. A Create Intent API records the instruction with constraints such as merchant, budget ceiling and product preferences; the customer signs off, and Amex issues an Intent ID plus a Proof of Intent Token that is exchanged for a scoped payment credential. The agent never touches the real card number, and the same Intent ID can be reused for disputes and fraud checks later. The design choice that matters is timing: verification happens before money moves, so a mismatch between "size 9 white Nike sneakers" and an ordered pair of Adidas gets blocked rather than arbitrated.

Google Finally Releases Gemini 4 Argon

Google ended months on the frontier sidelines with Gemini 4 Argon, its first major flagship since Gemini 3 in late 2025, now rolling out to Pro and Ultra subscribers and to developers through the Gemini API. The company claims Argon beats GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 across coding, cybersecurity, science, finance and law benchmarks, with a one-million-token limit and a continuation protocol for long responses. The launch price is the louder signal: $2 per million input tokens and $10 per million output, roughly a fifth of OpenAI's and Anthropic's frontier pricing, doubling after November 30. Early builder tests report comparable quality at a fraction of the cost, though Argon burns far more output tokens per task, so the headline discount may not survive contact with real workloads.

Character.AI Files for Bankruptcy and Sells Assets to Disney

Character.AI filed for Chapter 11 in the US Bankruptcy Court for the District of Delaware with assets and liabilities between $50 million and $100 million, and agreed to sell part of itself to Disney for $20 million as a stalking horse bid. The five-year-old startup raised over $193 million and logged 20 million monthly active users by 2024, after Google paid $2.7 billion to license its models. It never converted engagement into cash. In its filing the company said AI entertainment "has proven to be costly to develop and grow," with inference costs scaling linearly with usage and legal liability from a teen-safety lawsuit layered on top. It lands days after the Firmus IPO collapse reset AI funding expectations.

JPMorgan Draws the Liability Line on AI Agent Account Access

On JPMorgan's third-quarter earnings call, Jamie Dimon said the largest US bank will not let outside AI agents inside the company: "You don't want rogue agents, and it could be thousands of them." A spokeswoman confirmed the bank blocks third parties from obtaining customer banking credentials and logging in on their behalf; agents may read balances through open-banking connectors such as Plaid, but cannot move money or pay a bill. Buried in the same call was the line that should worry every company building on agents: "If people are using them, they should understand the risk. They should understand there's a liability transfer."

---

What to Watch Tomorrow

• Apple's HomeView and Siri AI: Apple launched a $299 smart home hub built to showcase AI-upgraded Siri, alongside a new Apple TV 4K and HomePod mini (CNBC).
• Independent evaluator control: A Reuters investigation reports Anthropic and OpenAI sought significant editorial control over what outside evaluators can publish; both labs dispute parts of the account (Reuters via StreetInsider).
• Memory stocks wobble: Micron, SK Hynix and other memory names fell for a second day after Chinese rival CXMT touted progress on more efficient memory manufacturing, with semiconductor ETFs on their longest losing streak since May 2025 (CNBC).

Builder Impact

The authorization layer is becoming the product. Amex's Intent ID and JPMorgan's refusal to accept credentials point at the same missing primitive: scoped, intent-bound, revocable credentials plus an audit trail that attributes an action to a specific agent. Handing an agent a raw API key or a stored card number is now a model the two largest payment players have publicly declared unsafe. Design for delegated credentials and agent-attributable logs.

Cheap frontier tokens will reprice your stack. Gemini 4 Argon undercuts OpenAI and Anthropic roughly five-to-one at launch, but it spends far more output tokens per task. Benchmark on cost-per-completed-task, not cost-per-token, and keep the abstraction layer thin enough to swap models before prices double on November 30.

The self-replication report is an infrastructure warning, not a science-fiction one. The capability Anthropic measured comes from the same coding and tool-use skills your agents already have. If your edge fleet, CI runners or container hosts use default credentials and open egress, close the egress path and rotate baseline credentials before the capability floor arrives.