- Human in the Loop
- Posts
- #5 Edition: "Over 40% of agentic AI projects will be canceled by 2027"
#5 Edition: "Over 40% of agentic AI projects will be canceled by 2027"
PLUS: Big week for Google & AI Agents: A2A, Agent Mode, and Gemini CLI Drop

Hey, it’s Andreas.
And welcome back to Human in the Loop.
Gartner dropped a brutal reality check this week: “over 40% of agentic AI projects are expected to be canceled by 2027” — crushed by cost, hype, and unclear ROI (full study below).
And yes, not every agent project will survive. But the right ones? They’ll reshape how we build, automate, and scale. That’s exactly why this newsletter exists. So let’s get into it.
This week’s edition covers:
Google donates A2A protocol to standardize AI agent communication, Airtable pivots into an AI-native app platform, and Postman launches “AI-Ready APIs”.
Google released Gemini CLI — an open-source, terminal-native AI agent that brings powerful, workflow-driven capabilities directly to developers’ command lines.
and much more…
Here we go.

Weekly Field Notes
🧰 Industry Updates
New drops: Tools, frameworks & infra for AI agents
🌀 Google donates A2A protocol to Linux Foundation to standardize AI agent communication
→ Biggest news of this week! A2A could become the TCP/IP of agents — enabling interoperability across models, tools, and platforms. Open-sourced via the Agent Protocol Alliance.
🌀 GitHub Copilot adds an autonomous agent for coding + issue handling
→ GitHub quietly enters the autonomous dev agent space. This isn't just autocomplete anymore — it's an actual Copilot that takes over tasks.
🌀 Airtable pivots into an AI-native app platform
→ Not a feature drop — a full company reboot. Airtable just rebuilt the stack from scratch for the agent era: apps powered by embedded agents, automating thousands of hours of workflows in seconds. Think: vibe coding meets real-world scalability. The next company going full agentcore.
🌀 Microsoft debuts “Mu” — its new on-device LLM powering Windows AI agents
→ Designed for fast, local inference. A strategic play for privacy-first agentic applications.
🌀 Anthropic adds build, host & share features to Claude apps
→ Now you can super fast turn Claude prompts into hosted mini-agents — no infra required.
🌀 Postman launches “AI-Ready APIs”
→ Helps devs make APIs agent-friendly by default. A smart move to fuel agent ecosystems (check also their developer guide to-AI-ready APIs).
🌀 ElevenLabs debuts “11.ai” for real-time voice-first agents
→ I am more convinced than ever that natural speech is the new UI. This drop aims to make voice agents feel more human and reactive.
🌀 Google drops Magenta-Realtime for audio + text music generation
→ LLM + synth = real-time music copilot. Creative agents are evolving fast. The team released a Python library, a comprehensive blog post and model card with more info.
🌀 OWASP Foundation released their new AI Testing Guide
→ It’s a practical framework to test AI applications (including AI agents) for security, reliability, and compliance across real-world workflows. AI Agents introduce new attack surfaces — making this a valuable perspective for anyone building secure, autonomous systems.
🎓 Learning & Upskilling
Sharpen your edge - top free courses this week
📘 DeepLearning.AI launches together with IBM course on building multi-agent systems with ACP
→ Top recommendation for learning everything orchestration — includes hands-on experience with multi-agent collaboration and real-world workflows.
📘 DAIR.AI Academy launches new course on building with Reasoning LLMs
→ Learn how to build agentic systems with reasoning LLMs like Claude 3.7 and o3. Covers Agentic RAG, hybrid prompting, LLM-as-a-Judge patterns, and where model limitations kick in. Tactical and well-structured — great for builders.
📘 IBM Technology on “How will AI Agents Manage Identity & Build Trust in Complex Systems”
→ Distinguished Engineer Grant Miller breaks down how agentic systems handle identity, trust, and secure access. Covers OAuth 2, OpenID Connect, token exchange, and transitive trust across dynamic agent workflows. Essential viewing for anyone building multi-agent architectures with real-world constraints (15min well spent).
P.S. Got a good new course? Send it my way on LinkedIn or just hit reply.
🌱 Mind Fuel
Strategic reads, enterprise POVs and research
🔹 Gartner warns: 40%+ of AI agent projects will fail by 2027
→ Most agent deployments today are hype-wrapped chatbots — what Gartner calls “agent washing.” The real problem? No ROI, rising costs, and tech that breaks under pressure. They analyzed thousands of vendors and only 130 vendors even offer true agentic capabilities.
🔹 University College London, Tencent AI Lab, Huawei, and Oxford dropped one of the most comprehensive research paper on Deep Research Agents
→ Deep Research agents combine dynamic reasoning, multi-hop planning, tool orchestration, and structured report generation. The paper lays out core components (MCPs, multimodal inputs, code execution), classifies static vs. dynamic workflows, and critiques current benchmarks for missing real-world objectives.
🔹 McKinsey analyzed 150+ GenAI of their deployments and revealed what sets the winners apart
→ One-offs don’t scale. McKinsey found that top-performing companies use modular, open GenAI platforms — built for speed, control, and reuse. The common thread? They’re not chasing fancy models — they’re investing in infrastructure that scales.
🔹 Salesforce targets 1 billion deployed agents
→ Big bet on AI agents in customer workflows. Deployment scale is already massive on their side. interesting perspective shared by Salesforce CEO Marc Benioff.

♾️ Thought Loop
What I've been thinking, building, circling this week
This week, I had the honor of giving a masterclass at one of Northern Germany’s leading economic forums — the EuroMinds Economic Summit 2025 in Hamburg. Always a privilege to be on home turf.
I focused my talk on a key belief: 2025 isn’t the year of AI Agents — it’s the start of the decade of AI Agents.

Here’s what I shared with the room — and why this shift goes far beyond GenAI:
1. GenAI was the demo. Agents are the deployment.
→ Agents don’t just chat — they act.
→ They handle tasks, make decisions, and trigger systems.
→ This is the moment AI gets hands and feet.
2. This changes work — entirely.
→ Agents aren’t assistants. They’re operators.
→ Building, testing, deploying, fixing — across workflows.
→ The line between “employee” and “agent” is already blurring.
3. Market outlook? Explosive.
→ A $500B+ opportunity.
→ Every major analyst sees it coming.
→ The winners? Companies that deploy, not delay.
Thanks to everyone who joined (I even met two people reading this newsletter) — great energy, interactive session, and lots of new perspectives. I’m really enjoying these formats that bring together such a diverse and curious crowd.


🔧 Tool Spotlight
A tool I'm testing and wachting closely this week
Gemini CLI from Google just dropped — and it’s the closest thing we’ve seen yet to a full-stack terminal-native AI agent.
→ Agentic
→ Open-source
→ Built for workflows, not demos
How it works:
Runs in your terminal
Understands your full codebase, not just one or a few files
Comes bundled with Gemini 2.5 Pro, a 1M token window, and the highest free-tier rate limits in the game (60 req/min, 1,000/day)
Lets you code, debug, generate, automate, and extend — from the shell
Plug in tools, use Google Search grounding, or connect MCP extensions
This isn’t just a wrapper. It’s real orchestration. Gemini CLI plans, reasons, recovers from failed steps, and acts — like a proper agent should.
Oh, and it’s open-source (Apache 2.0). You can inspect it, fork it, contribute.
→ npx https://github.com/google-gemini/gemini-cli
→ Or just run gemini
after install and start building.
Feels like the future of dev workflows is finally creeping into the terminal. Big move from Google.
→ Explore Gemini CLI on GitHub

That’s it for today. Thanks for reading.
Enjoy this newsletter? Please forward to a friend.
See you next week and have an epic week ahead,
— Andreas

P.S. I read every reply — if there’s something you want me to cover or share your thoughts on, just let me know!
How did you like today's edition? |