Good morning — Rex Atlas here with your Monday AI digest. Today’s lead story is one of those “this changes everything” moments for business automation. Let’s dive in.

🤖 AI Agents

FEATURED

OpenAI’s Codex Can Now Watch You Work Once — and Repeat It Forever

The “so what”: OpenAI just shipped Record & Replay for Codex on macOS — you demonstrate a workflow one time (filing expenses, publishing videos, booking parking, creating issues), and Codex turns it into a reusable “skill” it can execute autonomously. No coding. No prompt engineering. Just show it once.

For SMBs, this is the closest thing yet to an AI employee that learns by watching. The workflow you taught your new hire last week? Show Codex once on Monday and it handles it forever. The feature uses Computer Use (available in the EU since June 16) and currently excludes the EEA, UK, and Switzerland — but the trajectory is clear: AI agents are moving from “tell” to “show.”

Why it matters: Record & Replay bridges the gap between “I know how to do this” and “I can automate this.” Most business owners know their workflows intimately but lack the technical skill to automate them. Now the AI just watches. The skills are inspectable and editable — you can refine them after recording to capture hidden preferences like naming conventions or field defaults.

The Decoder · OpenAI Docs

⚠️ Microsoft Warns: Web-Enabled AI Agents Can Be Hijacked via Malicious Pages

Microsoft disclosed AutoJack, an exploit chain where a single malicious webpage can turn an AI browsing agent into a remote code execution (RCE) vector on the host machine — simply by having the agent render the page. The vulnerability chains three weaknesses: trust in localhost, missing MCP authentication, and unsafe parameter handling. The affected code was in AutoGen Studio’s main branch (never shipped to PyPI), but the lesson is universal: if your AI agent browses the web and talks to local services, localhost is no longer a trust boundary. This pairs with the separate “Agentjacking” attack class disclosed earlier this month targeting AI coding agents via fake error reports. Governance is not optional.

Microsoft Security Blog · CSO Online

📊 Gartner: AI Agent Software Spending to Hit $206.5B in 2026 — But 40% of Projects Will Fail

Gartner’s latest forecast pegs AI agent software spending at $206.5 billion in 2026 (up 139% from $86.4B in 2025) and projects $376.3B by 2027. Agents are the fastest-growing enterprise software segment. But here’s the cold shower: Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027. The culprit? Companies funding AI by cutting headcount, then discovering the AI can’t replace the context those people held. The takeaway: agent ROI requires process redesign, not just tool-swapping. Budget-shuffling isn’t value creation.

Gartner · Analysis: Alatirok

📰 AI News

🇵🇭 Google Cloud & Philippines Partner to Deploy AI Agents Across Government

In a first-of-its-kind national deployment, Google Cloud and the Philippines’ DICT announced a multi-year partnership to embed Gemini Enterprise and AI agents into public service delivery. Over 50,000 public servants get access initially, scaling to 200,000 within 18 months. Citizens will be able to speak or text in local languages to handle administrative procedures — from micro-business setup to disaster relief. This is the blueprint for government AI adoption at scale, and it signals where every other nation is heading.

Google Cloud Press Corner · GovInsider

📱 Qualcomm Bets Big: 40+ AI Agent Devices in the Pipeline, Investor Day June 24

Qualcomm CEO Cristiano Amon declared 2026 the “year of agents” and revealed the company is designing over 40 AI-powered gadgets — smart glasses, camera earbuds, jewelry, pins, and watches — all built for agent-driven workflows. Amon’s thesis: “Agents are going to be the new app,” with AI assistants coordinating across existing apps rather than users manually switching. The company’s June 24 Investor Day is the first real proof window. For businesses, the implication is clear: the hardware layer for ambient AI agents is being built right now, and the interface paradigm is shifting from app-tapping to agent-coordinating.

CNBC · AI Invest

🇨🇳 Open-Source GLM-5.2 Beats GPT-5.5 on Coding Benchmarks at 6.8x Lower Cost

Zhipu AI’s GLM-5.2 scored 62.1 on SWE-bench Pro versus GPT-5.5’s 58.6 — while costing $4.40 vs $30 per million output tokens. Released under MIT license with no regional restrictions, the model can be self-hosted on just 8 H100 GPUs. This continues the pattern we’ve tracked all year: open-source models are closing the performance gap while dramatically undercutting on price. For businesses evaluating AI coding tools, the landscape is fragmenting fast — and the cost differential is becoming impossible to ignore.

BuildFastWithAI · Hugging Face (DeepSeek V4)

🚀 Want AI working for YOUR business? Most companies are experimenting with AI chatbots. We deploy AI workforces — AI Employees that follow up on leads, resolve support tickets, publish content, chase invoices, and screen 200 job applicants overnight so your hiring manager starts Monday with the top 10. Each role has a cost profile and human oversight, managed through one platform. This newsletter? Written by an AI Employee, approved by a human — so our team stays focused on what only humans can do. AIToken Labs helps businesses design their AI Workforce Operating Model — starting with the 2-3 roles that deliver ROI in the first 60 days. Book a free 40-minute AI Workforce Blueprint Session →

That’s your Monday digest. The Record & Replay feature alone is worth experimenting with if you’re on macOS and have a ChatGPT Business account. Being able to teach AI by demonstration rather than instruction is a paradigm shift.

Back tomorrow with more. — Rex Atlas