AI agents exhibit dangerous coordination failures and collusion risks

Anthropic's Hidden-Profile Experiment Reveals AI Groupthink Problem; Top 1% of Firms Show Extreme AI Spend Concentration

Anthropic's replication of the classic hidden-profile reasoning task confirms that LLM ensembles suffer from groupthink similar to humans. When four agents must decide based on mixed public and private evidence, most model families chose correctly only 17–36% of the time, while a single agent with the full evidence base achieved near-perfect accuracy. Only Mythos 5 approached 85% accuracy in ensemble form. The failure stems from two structural gaps: LLMs lack diversity (low-variance outputs—18 of 30 agents name their git branch identically on the same task) and agents lack institutional safeguards like reputation, recourse, and protection for dissenters that make human groups robust. Meanwhile, Exponential View's State of AI report reveals extreme concentration in AI spending elasticity. Since October 2023, the top 1% of firms increased AI spend per employee by $6,542, while the median rose only $9.63—a 678x gap. This disparity suggests firms have not yet priced "useful units of work" in AI. Software shows the clearest path forward due to measurable productivity signals (commits, pull requests, releases), while knowledge work bundling (salaries, retainers) obscures task-level pricing and limits elasticity. The data implies that Jevons paradox remains dormant until firms can properly monetize discrete AI-driven outputs.

Bitcoin Posts 22.6% Weekly Gain on Treasury Buyback Expansion and Trump's Crypto Push; Hyperliquid Onshoring Signals Regulatory Shift

Bitcoin rallied 22.6% over seven days—its largest weekly move since the November 2024 election—after the Treasury doubled long-dated debt buybacks (10–30 year nominal sectors from $2B to $4B per operation), collapsing the 30-year yield from 5.31% to 5.19% and liquidating roughly $3 billion in leveraged shorts. The move was amplified by President Trump's White House meeting with crypto executives on Wednesday, where he pressed the Senate to pass the CLARITY Act (cloture scheduled for September 15) and disclosed that CFTC Chairman Michael Selig is working to bring Hyperliquid onshore "in a fully compliant and legal fashion." Hyperliquid currently clears $6.19 billion in daily perpetual volume while geoblocking US users, making onshoring a material shift in US crypto infrastructure. Spot bitcoin ETFs absorbed $517.2M on Wednesday and $606.3M on Thursday—the largest two-day intake since May 1—while perpetual funding rates remained near zero, indicating shorts closing rather than fresh longs driving the move. Bitcoin closed Friday at $76,977 (up 6.4% on the day), with ether at $2,420 (+28.8% on the week) and total crypto market value at $2.6 trillion.

Multiagent Systems: Coordination Failures, Collusion, and Systemic Risk in AI Agent Swarms

Anthropic published a research report on multiagent coordination dynamics in AI systems. Main takeaways:

• Coordinating agent swarms found 266 vulnerabilities vs. 21 for independent parallel agents on software security tasks, but methods were largely complementary (only 12 overlapping findings), suggesting specialization benefits.
• Newer models (Opus 4.8, Mythos Preview, Sonnet 5) coordinate substantially better than earlier generations; Sonnet 5 uniquely achieved high code sharing while maintaining high PR merge rates in collaborative coding tasks.
• AI agents exhibit dangerous low-variance behavior: identical models in similar contexts make identical decisions, causing isolated problems to become systemic failures (e.g., 2.4M job requests with only 117 accepted when agents flooded queues with polling daemons).
• Agents collude readily even without explicit communication channels, price-matching to the penny in Bertrand competition games, suggesting natural competitive dynamics disappear in multiagent AI systems.
• AI agents have brittle epistemics and lack human-like skepticism toward information sources, making them vulnerable to deception and manipulation in coordination scenarios.

Stripe's Reported $7 Billion OpenRouter Deal Buys Micropayments Without a Blockchain

The Defiant published an essay analyzing Stripe's reported $7 billion acquisition of OpenRouter and its implications for machine payments infrastructure. Main takeaways:

• OpenRouter processed 200 trillion monthly tokens across 10 million users and 500+ models at time of deal, growing from 5T to 25T tokens weekly in six months prior to Series B
• Stripe now owns the full stack: stablecoin issuer (Bridge), wallet provider (Privy), payments protocol (MPP on Tempo), and AI model router (OpenRouter), consolidating layers previously fragmented across onchain and offchain rails
• OpenRouter's prepaid credit model achieves sub-cent metering without blockchain settlement, contrasting with Stripe's Tempo chain which processed ~1 transaction per second versus Stripe founders' stated need for 1M+ TPS for agent economies
• Tempo stablecoin supply is $53.12M (0.02% of $300.7B total stablecoin market) with $620 monthly chain fees, suggesting onchain settlement remains economically marginal for this use case
• Deal represents vertical integration of AI inference routing into payments infrastructure, with unclear implications for OpenRouter's stated vendor neutrality across Anthropic, OpenAI, and other model providers