Hey, it’s Andreas.
Welcome back to Human in the Loop — your field guide to what’s next in AI agents, automation and real-world execution.

This week was huge for AI agents:

• AWS goes all-in with a $100M agentic AI innovation center, Capgemini forecasts a $450B agentic AI boom by 2028, and Uber drops Finch, a Slack-native finance agent.

• I’ll also walk you through a dead-simple guide to MCP — How to set up MCP and turn Claude into your AI ops layer — in 10 minutes or less.

• and much more…

Let’s get into it.

Weekly Field Notes

🧰 Industry Updates
New drops: Tools, frameworks & infra for AI agents

🌀 AWS backs Agentic AI with $100M Innovation Center
→ Announced this week at AWS Summit New York, alongside several major updates to AWS's AI agent capabilities: Amazon Bedrock launched AgentCore, a new “Agents & Tools” category was added to the AWS Marketplace; and Amazon Nova introduced custom model building capabilities.

🌀 OpenAI launches ChatGPT Agent
→ Pro users can now assign tasks, set goals, and let ChatGPT plan, browse, and execute autonomously — all inside the app.

🌀 Mistral releases Voxtral voice model
→ Open-source. 40+ minutes of audio. Built for agent pipelines.

🌀 LangChain drops Open Deep Research agent
→ Multi-agent chains for automated research reports.

🌀 Reflection launches Asimov — a code research agent for teams 
→ Built for deep codebase understanding: Ingests code, docs, and chat history to act as a shared engineering brain. Features team memory, RBAC, and a multi-agent architecture — performs 60–80% over other tools in blind tests. There is a waitlist for now.

🌀 Galileo drops Agent Leaderboard v2
→ The most comprehensive eval yet for enterprise agents — 5 domains, 100+ conversations each, tracked by Action Completion and Tool Selection Quality. Last results: GPT-4.1 leads overall, Gemini 2.5 shines on tools, Kimi K2 tops open source.

🌀 GitHub Copilot Agent gets a major update
→ Now tests UI with Playwright, logs screenshots in PRs, connects to external tools via MCP, adds a task dashboard — and only uses one Copilot request per session.

🌀 Uber launches Finch
→ Slack-based AI finance agent with real-time insights.

🌀 Cased releases interactive DevOps agents 
→ Always-on agents that handle the busywork: fix infra drift, spot cloud cost savings, and assist when deploys go wrong.

🎓 Learning & Upskilling
Sharpen your edge - top free courses this week

📘 OpenAI Academy workshop on PydanticAI agents (live in 7 days)
→ Build safe, structured agents using strong typing and validation.

📘 Google prompt strategies to improve agents
→ Proven patterns for reliable, safe, grounded agents.

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

🔹 Stripe on Agentic Commerce
→ On The MAD Podcast, Stripe’s Emily Sands explains how buying bots, MCP, and intent-based interfaces are reshaping transactions. Stripe’s own payments model boosts fraud recall to 97% and powers real-time risk scoring, tokenized payments, and AI-native billing. Interesting watch!

🔹 NVIDIA on Small Language Models are the future of agentic AI
→ New research shows SLMs (2–9B params) beat many large models on agent tasks. They're faster, cheaper, and better suited for scoped workflows. Key reasons: lower latency, better alignment, on-device deployment, and stack modularity.

🔹 IBM asks “Do you need a CAIO (Chief AI Officer)”?
→ New study across 2,300 companies: orgs with a CAIO see 10% higher ROI, yet only 26% have one. Key takeaways: centralized org models win, measurement is hard but essential, and CAIOs must own the budget.

🔹 Capgemini: Agentic AI to add $450B by 2028
→ Massive 100+ page report. Highlights include: Human-agent collaboration dominates, but trust is down to 22%. Only 14% deploy at scale, yet 93% see agents as future edge. And agent performance doubling every 213 days. Good read!

♾️ Thought Loop
What I've been thinking, building, circling this week

I’ve been in tech for 10 years — and MCP (Model Context Protocol) is one of those rare innovations that actually lives up to the hype. If you’re not using it yet, you’re leaving serious leverage on the table.

(Keep reading — I’ll show you how to get it running in under 10 minutes. Plus: the best MCP servers to start with.)

What is MCP (Model Context Protocol)?

Think of MCP like USB — but for AI tools. Released by Anthropic and built to work natively with Claude, MCP lets your AI interact with tools like Notion, Slack, HubSpot, Google Calendar, CloudWatch, and even your local files.

Instead of copy-pasting between tabs like it’s 1998 — just plug in MCP and let Claude handle the rest:

→ Post messages
→ Pull live data
→ Organize files
→ Orchestrate across apps
→ All in one go.

And the best part? One connection = access to an entire ecosystem of tools.

Yet many people get tripped up by the name. “Model Context Protocol” sounds like deep-tech sci-fi — but it’s not.

At its core, MCP is a simple standard that lets your AI talk to your tools in a language they all understand.

The best part? It’s incredibly easy to set up — no custom integrations, no complex APIs, no engineering background needed.

And the upside? Surprisingly high.
You can 10x the power of Claude — without writing a single line of code. Just connect, prompt, and let your AI do the work.

How to set up MCP (Claude version, 10-min max):

1. Download Claude Desktop
   (Web version doesn’t work yet)

2. Install Node.js
   → nodejs.org (just click install)

3. Install Docker Desktop
   → docker.com/products/docker-desktop (just click install)

4. Enable MCP Toolkit
   Inside Docker: Settings (bottom left) → Beta Features → Enable MCP Toolkit

5. Add your first MCP server
   Start with e.g. Notion. In Docker MCP UI, connect Notion → copy token → enable scopes → connect workspace

6. Tell Claude to use it
   Claude will now suggest tools from Notion directly inside the chat window.

Power Tip: Start simple

Start with a simple MCP like Notion or Apify. Feel the magic and then add more. MCP is still early. But you will realize very fast that the leverage it gives you is amazing.

There are now 1000s of MCP servers:
→ Web scrapers, CRMs, vector DBs, YouTube transcriptors, file ops etc.
→ You’ll need a bit of time to explore which tools work best for your workflow — ideally the ones you use most or where your key data lives.

📚 Want more?

→ Free 75-page MCP guidebook with 10+ build-ready examples:
📘 Cookbook link

→ Amazing list of 12 must-try MCP servers:
🧰 Toolbox link

🔧 Tool Spotlight
A tool I'm testing and watching closely this week

This was a big week for AWS with several major announcements around agentic AI (see above). But one AWS launch flew under the radar...

The launch of Kiro — a new IDE built from the ground up for agentic development.

I’ve used Claude Code for months — and it already crushes Copilot and Cursor.
It doesn’t just autocomplete. It architects. Plans. Reasons. It builds real features, not just snippets.

Kiro builds on the exact same paradigm: goal-oriented, spec-driven software creation.
And it’s exciting to see AWS stepping in to push this category forward. It runs locally, works cloud-agnostically, and supports MCP out of the box.

What makes it different?
Kiro introduces specs and hooks as first-class citizens in the development flow.

→ Specs are structured artifacts that guide the agent’s planning — they bring clarity, reduce guesswork, and make feature development reproducible.
→ Hooks trigger background automation, letting agents execute tasks as part of a larger workflow — not just code suggestions, but full orchestration.

Agent IDEs are becoming a serious category. Cursor, Continue, and now Kiro are redefining how we build software. And with AWS behind this, it’ll be interesting to see how aggressively they scale it — and whether Kiro becomes the blueprint for how agents enter real-world dev workflows.
→ Explore Kiro (currently free to use — and surprisingly easy to set up.)