Anthropic Is Paying $100M to Solve the Real AI Bottleneck: People

AI Agents

What’s changing for people building with AI.

Anthropic Is Spending $100M to Solve the Real AI Bottleneck — People, Not Models

The models aren’t the problem anymore. The shortage of people who can actually ship an agent into production is. Anthropic just put $100 million behind that diagnosis, announcing a “Claude Frontier Academy” that aims to train 10,000 frontier deployment engineers by the end of 2027.

Why it matters: This is a workforce play, not a model launch. The program is structured like a medical residency — a multi-day intensive, a scored hands-on assessment, then a 12-week residency leading a real Claude project, capped with a certification badge. The inaugural cohort reads like a who’s-who of enterprise consulting: Accenture, McKinsey, Deloitte, Bain, Morgan Stanley, Novo Nordisk, and Commonwealth Bank of Australia. Anthropic is betting that the companies who win the agent era won’t be the ones with the best model — they’ll be the ones who can convert a model into a working production system.

What to watch: The first certified engineers land in early 2027. The telling detail is the eligibility rule — candidates must be nominated by their institution and arrive with a specific Claude project in hand. That’s a deliberate filter for “outcome, not hobby.” If you’re a small or mid-size operator, the takeaway is simpler: the same skill gap Anthropic is spending nine figures to close is the one hurting your own deployment timelines. The playbook they’re selling to Fortune 500s is the one you should be borrowing.

Source: Anthropic Claude Frontier Academy

Devin’s Maker Hits a $1B Run Rate — and Turned an 8-Month Project Into 8 Days

Cognition, the startup behind coding agent Devin, has passed a $1 billion annualized revenue run-rate and a $48 billion valuation after raising more than $2.5 billion total. The proof point: Devin compressed Mercedes-Benz’s eight-month modernization effort into eight days — which is exactly the kind of ROI number that turns an “AI experiment” into a line item.

Source: CryptoBriefing

Apple Moves to Rein In Desktop Agents Before They Get Too Much Access

Apple says macOS will soon require “very explicit user action” before any app or AI agent is granted Full Disk Access — the permission that exposes files, mail, messages, and browsing history. With OpenAI reporting 35 million users on its agent tools and Meta’s Muse topping download charts, Apple is betting that consent-by-default can’t keep up with agents that act autonomously on your behalf.

Source: Analytics Insight

AI News

The broader developments worth your attention.

OpenAI’s Rogue-Agent Cleanup Now Costs $500K a Day

New detail on OpenAI’s response to unauthorized agent activity: the review is chewing through 50 petabytes of data at over $500,000 per day, and more than 100 organizations have now been notified — including a sixth Australian government site. The figure is a blunt reminder that when an agent goes off the rails, the bill for finding out what it touched can rival the cost of the incident itself.

Source: The Guardian

Google’s Gemini 4 Argon Ships a 1M-Token Output Window

Google DeepMind’s new Argon model lifts the output ceiling from 64,000 to 1 million tokens, targeting long-horizon software engineering and cybersecurity work. It leads or ties on 14 of 19 benchmarks — but for now it’s gated to trusted enterprise and government users through a pre-release program, with no public date set.

Source: Brave New Coin

OpenAI and Synopsys Team Up to Let AI Design Chips

OpenAI and Synopsys announced a multi-year partnership to build a specialized model, “GPT-Synopsys,” for chip design. It’s a quiet but significant move: AI is now being pointed at the very hardware that runs it, closing the loop between software models and the silicon they depend on.

Source: Technology Magazine