This is the news. Agent dossiers — when to pick one, when to skip it — are on the agents page. Practical guides support your own testing and deployment decisions.
Cohere dropped a 218B MoE open-weight model with Apache 2.0 licensing, tool use, multimodal, and 48-language support on May 20. Nobody's talking about it. They should be.
The week's most important AI, agent, and automation news — curated and analyzed through a builder's lens.
Every AI agent demo looks incredible. Here's what separates the agents that survive contact with production from the ones that fall apart the moment real users touch them.
On May 5, 2026, a startup called Subquadratic shipped a model with a 12 million token context window and a claim that it's not a transformer at all. That's not a small thing. That's the first commercial challenge to transformer architecture dominance in two years.
While the industry debates agent architectures in conference talks, OpenHands quietly shipped an open-source agent that can navigate a browser, write and execute code, and take actions in real environments. Here's why that's the most important thing happening in open-source AI right now.
Every week brings another announcement of a longer context window. GPT-5.5 Instant ships with 256K. Gemini goes to 2M. SubQ hits 12M. Meanwhile, the actual problem — that models don't reason better with more context — gets ignored because it doesn't fit the marketing narrative.
If you're still fighting with inconsistent AI output formats, you're using the wrong approach. Structured output via JSON schema — done properly — eliminates the entire class of parsing problems that make AI pipelines fragile. Here's the exact pattern.
50%+ of Fortune 500 developers use AI coding agents weekly. The IDE is now an AI-first interface. Here's why that matters for every engineering team that hasn't made the switch yet.
Every engineering manager celebrating their team AI tool adoption is quietly watching their juniors become dependent on autocomplete. The productivity spike is real. The skill atrophy is worse.
While everyone was arguing about benchmark scores, Subquadratic shipped subquadratic sparse attention — and it changes everything about what's possible with long-horizon AI agents.
Retrieval-Augmented Generation became the default architecture for LLM applications because developers couldn't trust context windows. That says more about the industry than it does about the technique.
While the agent framework wars rage on, Composio quietly shipped a CLI that solves the actual pain point: discovering, configuring, and managing tools for AI agents. Here's why it matters.
The UK AI Security Institute proved frontier AI can autonomously run end-to-end offensive cyber operations. Here's what that means for every builder working with AI agents today.
HBR argues AI is making teams faster than their managers can keep up. Jeff Gothelf reads Karpathy as declaring vibe coding obsolete and replacing it with product management. The underrated skill of 2026 is still handling conflict well.
Anthropic's Mythos model is forcing the ECB to convene European banks around a frontier AI that writes zero-day exploits on first try 83% of the time. The neocloud market is heading to $400B by 2031. And the Model Context Protocol is going stateless with formal long-running-tasks support.
Anthropic shipped ten finance-agent templates on May 5 — reconciliations, KYC, month-end close, the boring work — and within ten days had PwC committing 30,000 certifications and KPMG wiring Claude into Digital Gateway. Goldman, JPMorgan, Citi, AIG and Visa are already running it. OpenAI answered with its own PwC finance deal a day later. The question for every fintech app is now structural: are you the recipe, or the pantry?
CVE-2026-48172 is a CVSS 10.0 LiteSpeed cPanel flaw being actively exploited — CISA KEV, May 29 deadline. The first public Apple M5 kernel exploit took five days with help from Claude Mythos, bypassing (not breaking) MIE. And the Mini Shai-Hulud worm compromised 160+ npm and PyPI packages by staging throwaway accounts and abusing CI tokens. Default installs are no longer safe.
Cisco's 2026 WAN study says agentic AI will push enterprise traffic from 2.5x to 9x by 2035, with AI inference flows making up 25% of all traffic and lasting 2x longer than web flows. Google Cloud's COO publicly disclosed that attack chains now move from breach to handoff in 22 seconds and that a single Maps API key silently gained Gemini access and drained $10,138 in 30 minutes. Credential brokering is the new minimum for least privilege.
General-purpose AI agents are displacing specialized software at the application layer, with value migrating to data and trust. The new 'AI services' stack is built around Action (orchestration) and long-running task management, with checkpoints, resume, identity propagation and scoped credentials as first-class requirements. Hugging Face has standardized the vocabulary. The 'long-running agent' is the new primitive.
Apple's 20th-anniversary iPhone for 2027 is shaping up to be the biggest industrial-design reset of the decade, with curved glass, under-display Face ID and solid-state buttons. Google Stitch now does real-time vibe design with voice and text, generating 5 UI screens at once on a Gemini-3 infinite canvas. Replit Agent 4 ships Custom Instructions, Skills, a Databricks integration and a Package Firewall that blocks 8,000 malicious packages a day.
SEC Commissioner Hester Peirce confirmed on May 22 that the innovation exemption for onchain stock trading will cover only tokenized NMS stocks — narrow, conservative, real. The tokenized RWA market has crossed $34 billion, up 10x in under two years, with tokenized U.S. Treasurys and gold making up two-thirds. The composability gap is the structural question: most tokenized assets are digital receipts, not DeFi building blocks. AI agents settled $73M across 176M onchain transactions in 12 months.
Microsoft is cancelling most direct Claude Code licences inside its Experiences and Devices group, telling engineers to migrate to GitHub Copilot CLI by 30 June 2026 — the most credible market signal that token-priced AI coding is not yet a sustainable enterprise line item. Uber burned through its entire 2026 AI coding budget in four months. Nolan Lawson's contrarian read: use AI as a senior reviewer, not as an autocomplete, and write high-quality code more slowly. Google's 'Modern Web Guidance' is a free, 100+ use-case developer reference.
Google is pushing AI Mode to a billion users and is now testing ads in AI Mode via the Universal Commerce Protocol — Ulta Beauty is the flagship. LinkedIn has started suppressing low-quality AI-generated content, with a 'low-quality AI' reporting category and behavioral-signal detection catching roughly 94% of obvious AI slop. The 2026 influencer marketing playbook is now 10–25% of total marketing budget, with 50–60% of that going to nano and micro creators, and TikTok the clear focal point.
OpenAI shipped GPT-5.5 Instant to every ChatGPT user on May 5, 2026 — replacing the default model for hundreds of millions of people without anyone noticing. That's the story. Not the benchmarks. The scale.