Global AI Weekly - Issue 172 - Global AI Community

In this issue

Highlights

2 stories

Accenture's AI Reality Check: $1,000 Claude Bills Meet a $29 Budget

Accenture spent months encouraging employees to embrace AI. Now the consulting giant is tightening the purse strings. After some staff reportedly racked up Claude bills approaching $1,000 per month, the company is limiting centrally funded usage to just $29 per employee. Additional consumption must be charged to project budgets. It's a revealing moment for enterprise AI adoption: enthusiasm is easy, but controlling token costs and proving business value is becoming the real challenge.

semafor.com

Perplexity Just Open-Sourced Its AI Agent Security Sandbox

AI agents are powerful, but what stops them from going rogue? Perplexity has open-sourced SPACE, the sandbox technology behind its Computer agent, giving developers a way to isolate potentially dangerous agent activity. The company claims nine frontier models failed to escape its sandbox across 118 attempts. With CoreWeave, Akamai, and Crusoe supporting adoption, SPACE could become an important building block for safer autonomous AI systems.

In this issue

Research

2 stories

arxiv.org

Harness²: What If AI Agents Could Improve Themselves Without Retraining?

What if better AI agents didn't require better models? Google's Harness² introduces a recursive self-improvement framework that lets agents optimize their own instructions, tools, and workflows without needing ground-truth answers. By testing changes on individual tasks and learning from previous improvements, the system continuously refines how agents operate. Across three professional benchmarks, researchers report performance gains of up to 11.8 percentage points. The takeaway: smarter agent infrastructure might matter as much as smarter models.

arxiv.org

Can AI Learn New Facts Just by Learning to Draw Them?

What if generating images could actually make language models smarter? Researchers investigate whether multimodal AI can acquire knowledge through visual generation and transfer it to text-based reasoning. Their Knowledge Transfer Probing framework tests this connection across seven knowledge categories and three models. The results show that information learned visually can influence language understanding, although inconsistently. It's an intriguing step toward AI systems that learn across modalities rather than treating images and text as separate worlds.

In this issue

Video

2 stories

youtube.com

Every Generation Thinks This Time Is Different: The AI Job Myth

very generation fears that new technology will destroy jobs. From automobiles replacing horses to ATMs transforming banking, history tells a more complicated story. In this video, IBM's Jeff Crume and Bri Kopecki explore why automation often changes jobs rather than eliminating them entirely. They explain Jevons Paradox, how efficiency can increase demand, and why AI might create opportunities we cannot yet imagine. Is this technological revolution really different, or have we heard it all before?

youtube.com

Your AI Coding Agent Is Confidently Wrong. Here's How to Fix That.

AI coding agents can write code at incredible speed, but they don't always recognize their own mistakes. In this AgentCon talk, Rob Zuber, CTO of CircleCI, explains why reliable agentic development depends on better feedback loops, not just smarter models. Discover how stop hooks, automated reviews, and remote validation can catch problems before they reach production. The goal is simple: help agents fail fast, fix mistakes early, and ship code developers can actually trust.

In this issue

Articles

2 stories

devblogs.microsoft.com

Azure MCP Server 3.0 stable release: More tools. More possibilities

Microsoft has released Azure MCP Server 3.0, giving AI agents a much bigger toolkit for managing cloud environments. New integrations cover Azure Backup, resilience management, SRE investigations, monitoring, and Cosmos DB, while improved tool schemas make integrations more predictable. Security-focused defaults and better authentication support help teams operate more safely. With over 50 beta iterations behind it, this stable release brings practical Azure operations directly into GitHub Copilot and other MCP-compatible agents.

explainx.ai

Your Kid's Next Playdate Could Be an AI Spy Mission

Think corporate espionage happens only in dark alleys? Think again. OpenAI reportedly trains employees to recognize potential spies in surprisingly ordinary situations, including conversations with other parents at playgrounds. Similar warnings have reached employees at Anthropic and Google. As competition for AI breakthroughs intensifies, protecting proprietary research now extends beyond office networks and cybersecurity. It's a fascinating glimpse into how the AI race is changing workplace security, and even everyday social interactions.

In this issue

Upcoming Events

2 stories

twilio.com

Nov 3 - Twilio Assemble London: The Future of Comms + AI

How do you get AI agents out of the UI and into the channels your customers already use? Join Twilio for an evening exploring the future of AI and communications, with practical examples, a look at Twilio Conversations, and TwilioWorld: Escape the IVR — a hands-on build challenge where you’ll put your skills to the test and compete for prizes. Expect useful ideas, plenty of building, good food and drinks, and a room full of developers, founders and tech leads.

neo4j.com

NODES 2026: 24 Hours of Graphs, AI Agents, and Smarter Applications

What happens when AI meets connected data? Join Neo4j's NODES 2026, a free 24-hour virtual conference bringing together developers, architects, and AI engineers from around the world. Explore GraphRAG, knowledge graphs, multi-agent systems, and modern application development through technical sessions, lightning talks, and expert discussions. Whether you're building smarter AI agents or exploring graph-powered applications, NODES offers practical insights to help turn connected data into real-world intelligence.

In this issue

Code

2 stories

github.com

Turn Your Favorite Chat App Into an AI Agent With MuseBot

Why build another chatbot when you can bring AI into the apps you already use? MuseBot is an open-source project that connects powerful language models to Telegram, Discord, Slack, WeChat, and more. Supporting OpenAI, Gemini, DeepSeek, and OpenRouter, it offers streaming responses, image and voice interactions, MCP tool calling, RAG, and scheduled tasks. Deploy it locally or with Docker and turn everyday conversations into AI-powered workflows without building everything from scratch.

github.blog

GitHub Copilot HydraFusion: Why Settle for One AI Model When You Can Have a Team?

Why choose one AI model when GitHub Copilot can orchestrate several? HydraFusion combines models from different providers to draft, review, and improve code through intelligent workflows. Its latest update introduces HydraFusion Max, parallel AI reviewers, and smarter reasoning allocation, using models including GPT-6 and Claude Opus 5.5. Available in research preview across VS Code, Copilot CLI, and the Copilot app, it explores a new approach to coding: let models collaborate instead of compete.

In this issue

Podcast

1 story

open.spotify.com

Ep #21: Cassie Groos (Unily): Design Systems for Agents, Infinite Brands, and Why the Canvas Stays

What happens when AI agents become both the builders and consumers of your design system? In this episode of Mostly Working, hosts Tom Harman and Jonny Burch talk with Unily's Cassie Groos about agent-ready design systems, structured component documentation, and managing design tokens across countless brands. From MCP servers to Claude-assisted workflows, the conversation explores how designers can keep control while AI handles more implementation. A practical look at where design meets agentic development.