The AI agent boom is hitting a cost wall. Microsoft is canceling thousands of internal Claude Code licenses after API bills exceeded the salaries the tools were meant to augment, with June 30 as the cutoff for its Experiences + Devices team (Windows, M365, Outlook, Teams, Surface). Gartner doubled down on its prediction that over 40% of agentic AI projects will be canceled by 2027 due to escalating costs and unclear business value. Meanwhile, enterprise automation is surging: n8n hit a $5.2B valuation after SAP took a stake, embedding workflow automation into its Joule Studio. Google Cloud's 2026 Agent Trends report confirms the shift from chatbots to autonomous workflow orchestration. In crypto+AI, stablecoins are emerging as the default settlement layer for autonomous agents, and Tether is launching a Georgian Lari stablecoin. The arc of 2026 is becoming clear: the agents that survive will be the ones delivering measurable ROI, not demos.
Microsoft's Experiences + Devices team is winding down Claude Code usage by June 30, 2026, after API bills surpassed what the tools were expected to save in developer productivity. Engineers are being pushed toward GitHub Copilot CLI. The move follows a broader pattern: Uber burned through its full-year AI budget by April, and multiple enterprises are reassessing agentic AI costs. cryptobriefing.com
Escalating costs, unclear business value, and inadequate risk controls will kill more than 40% of agentic AI projects by 2027, according to Gartner. The prediction is gaining traction as real cost data pours in from enterprises — Microsoft's Claude Code pullback being the latest data point.
Berlin-based n8n, an open-source workflow automation platform, hit $5.2B valuation after SAP took a secondary stake and embedded n8n into Joule Studio. A rare DACH-to-DACH agentic AI play at enterprise scale, reshaping how German banks and insurers deploy AI agents.
Google Cloud's 2026 report declares the "agent leap" — AI orchestrating complex, end-to-end workflows semi-autonomously. For enterprises struggling with speed-to-value, this is the defining opportunity of 2026: deterministic guardrails, context engineering, and headless CRM access are the new battlegrounds. cloud.google.com
Keyrock's new report finds stablecoins becoming the default payment rail for AI agents, handling sub-dollar transactions more efficiently than card networks. Meanwhile, Tether and the Georgian government announced GEL₮, a Georgian Lari stablecoin under the country's new digital asset framework. coindesk.com
03 / In their own feeds
Ten voices on X.
Selected posts from the edition, with links to the originals.
The AI Colony R&D · @TheAIColonyRD01 / X
"Microsoft just banned 100,000 engineers from using Claude Code. This wasn't because it doesn't work, but because the API bill was literally higher than the salaries it was meant to replace. Between this and uber burning their whole annual AI budget by april, the 'AI will save money' thesis is facing a reality check."
"When you're evaluating an AI agent, when can you get away with vibe evals, just eyeballing the outputs and trusting your gut? And when do you have to do the painful version: axial coding, where a human annotates every output?"
"🇺🇸 Utah just approved an AI data center twice the size of Manhattan. It's drawing 9 gigawatts of power and 16.6 billion gallons of water a year in a state already facing severe drought conditions. We'll see if Utah can make the numbers work."
"Companies incorrectly reported as stopped using AI: Microsoft: expanding AI spending, not reducing it. IBM: continues heavy AI investment while restructuring certain roles. Salesforce: increasing AI deployment despite workforce changes. Cisco: investing heavily in AI while..."
Security Disclosure: "OpenClaw shipped a bug that let strangers control your AI agent. If you ran the synology-chat plugin in versions 2026.2.22 or 2026.2.23 with an empty user allowlist — which is what you'd have if you set up the integration but hadn't added anyone yet — authorization was bypassed entirely."
"Language-only AI is trapped in ambiguity. It can produce brilliance and garbage in the same breath because ordinary language has no hard execution boundary. Formal math provides that boundary — turning knowledge into an environment where AI can safely evolve. That is the real break."
Qwen3.7-Max: "阿里通义千问发布 Qwen3.7-Max，专为 Agent 打造。亮点：35小时自主完成1000+工具调用，10倍性能提升；支持 Claude Code/OpenClaw 等多框架；AI Index 达56.6分。中国模型在长时自主任务上已达国际一流。" (Translation: Alibaba's Qwen3.7-Max is purpose-built for agents — 35 hours autonomous operation, 1,000+ tool calls, 10x performance boost. Supports Claude Code/OpenClaw frameworks. Chinese models now competitive on long-duration autonomous tasks.)