Google researchers unlock some truths about getting AI agents to actually work

Adoption of AI agents remains limited and uneven, with a McKinsey survey showing a majority of firms have yet to adopt agents, 40% experimenting, and under 25% deploying at scale (no function saw more than ~10% fully scaled or scaling; IT led with 2% fully scaled and 8% scaling), largely because multi‑step workflows are error‑prone and expensive. New Google research—180 controlled experiments across Google, OpenAI and Anthropic models—provides a practical rubric: single agents outperform multi‑agent setups on sequential tasks if accuracy exceeds ~45% (multi‑agent configurations cut performance by 39–70% there), but centralized multi‑agent architectures excel on parallel tasks such as finance (about +80% vs a single agent; independent parallel agents about +57%), guiding where firms should invest to achieve ROI. These technical findings arrive amid market and policy noise—President Trump’s executive order to pre‑empt state AI rules (likely to draw legal challenges), Oracle stock volatility tied to data‑center reports and $248bn of lease obligations, OpenAI’s faster image model and senior hires, Google DeepMind’s SIMA2 advances, and an FLI safety index putting top labs in the C range—highlighting regulatory, balance‑sheet and safety risks investors must factor into AI strategies.

Analysis

Adoption of AI agents remains limited and uneven: a McKinsey survey cited in the article found a majority of firms have yet to adopt agents, 40% are experimenting, and under 25% have deployed agents at scale; no functional area reported more than ~10% fully scaled or scaling (IT led with 2% fully scaled and 8% scaling). The article attributes slow, narrow adoption to engineering challenges—multi-step workflows amplify early errors—and to the high inference costs of capable models, which together raise the bar for positive ROI.

Google’s controlled study of 180 experiments across Google, OpenAI and Anthropic models offers actionable configuration guidance: single agents outperform multi-agent setups on sequential tasks if a single-agent accuracy exceeds roughly 45% (multi-agent designs reduced performance by 39–70% on those tasks), whereas centralized multi-agent architectures excel on parallel tasks (financial-analysis benchmark: centralized multi-agent ~+80% vs single agent; independent parallel agents ~+57%). These results imply task-specific investment priorities rather than one-size-fits-all platform bets.

Market and policy signals increase near-term risk: Oracle shares were hit after data-center delay and a reported $248bn of lease obligations (Oracle disputed delays), President Trump signed an executive order enabling federal challenges to state AI rules (likely litigation), OpenAI launched a 4x faster image model and made senior hires, Google DeepMind released SIMA2, and FLI’s safety index keeps top labs in the C range, underscoring regulatory, balance-sheet and reputational risks investors must price into AI exposure.

Key Decisions for Investors

• Reduce or hedge near-term exposure to Oracle (ORCL) until data-center execution and the $248bn lease-payment profile are clarified,
• Favor investments in firms able to demonstrate task-specific agent ROI—prioritize vendors and customers where parallel finance/analytics workflows can use centralized multi-agent architectures (per Google research) and show measurable performance gains,
• Require proof points before backing broad agent rollouts: look for single-agent accuracy >45% for sequential workflows or cost-effective centralized multi-agent implementations for parallel tasks,
• Monitor regulatory risk and litigation stemming from the Trump executive order and state responses; factor potential compliance costs and federal funding threats into valuation scenarios,
• Incorporate AI safety momentum into conviction: prefer labs with improving FLI scores (OpenAI, Anthropic, Google DeepMind) and watch for deteriorating "current harms" metrics that could invite regulatory scrutiny