Hey, it’s Andreas.
Welcome back to Human in the Loop — your field guide to what just dropped in AI agents, and what’s coming next.

From Google turning “Nano Banana” into a serious Photoshop rival … to Microsoft building its own LLM stack … to NASA + IBM releasing a new model — another exciting week.

But the big story of the week?
A landmark Stanford study showing how AI is already disrupting entry-level jobs — the first clear evidence of where the impact lands first.

Let’s dive in.

Weekly Field Notes

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

🌀 Google’s Nano Banana revealed as Gemini 2.5 Flash Image model
→ Turns out “Nano Banana” isn’t a meme — it’s Google’s new Flash Image model inside Gemini 2.5. Prioritizes fast, consistent edits and brand fidelity, setting it up as a real Adobe Photoshop competitor.

🌀 Microsoft drops MAI-Voice-1 + in-house LLM MAI-1
→ Microsoft has launched its first proprietary AI models, MAI-Voice-1 and MAI-1-preview, marking a strategic pivot toward building its own AI stack (not just leveraging OpenAI).

🌀 IBM + NASA unveil Surya solar weather model
→ Trained on 9 years of high-res solar data, forecasts flares 2 hours early with 16% higher accuracy. First to visually predict eruption zones. Open-sourced on Hugging Face — AI for good protecting satellites, energy, and connectivity.

🌀 OpenAI launches gpt-realtime API for next-gen voice agents
→ Enables developers to build truly reactive, voice-native agents. Low-latency, streaming, and multimodal.

🌀 IBM’s ACP joins Google A2A under Linux Foundation
→ IBM’s Agent Coordination Protocol now aligned with Google’s A2A standard — consolidation around open multi-agent infra is accelerating.

🌀 xAI debuts Grok Code Fast 1 with 256k context window
→ Focused on coding speed + context. A 256k window means whole repos can sit in context. Another clear signal: xAI wants a seat at the dev-agent table.

🌀 Mistral AI releases Medium 3 with stronger reasoning & vision
→ A mid-size model, but sharper multi-hop reasoning and vision handling. Mistral keeps proving smaller models can still punch above their weight.

🌀 Alibaba open-sources AgentScope for multi-agent apps
→ AgentScope provides a framework for coordinating multi-agent teams. Open-source release adds weight to the fast-growing Chinese open agent stack.

🌀 Netflix builds multimodal Media Data Lake with LanceDB
→ Not a model drop, but infrastructure: Netflix is building a data lake optimized for text + video + audio. A foundation for agent-native content search and production.

🌀 Anthropic pilots Claude for Chrome
→ A light extension for in-browser help. Looks simple, but could bring Claude into everyday workflows far beyond chat.

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

📘 Microsoft unveils Copilot Agent Academy
→ A new hub for devs building Copilot agents. Think structured curriculum + templates to accelerate enterprise agent adoption.

📘 Anthropic drops 8 guides on working with AI systems
→ Covers everything from prompt engineering to agent design, coding with Claude, and scaling AI adoption. Tools will change — fundamentals won’t. Some of the best tactical material to study right now.

📘 DeepLearning.AI (+ Neo4j) launches course on Agentic Knowledge Graphs
→ Timely release. Shows how to combine agents with structured knowledge bases — critical for grounded, reliable reasoning.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 OpenAI’s Shunyu Yao: “We’re at AI’s halftime”
→ The first half of AI was about building methods and models (from DeepBlue to GPT-4). The second half, starting now, shifts from training to evaluation — from “can we solve X?” to “what should we solve, and how do we measure progress?” Yao argues RL + reasoning has given us a general recipe for intelligence, so the frontier moves to defining tasks, evaluation, and real-world utility. A mindset shift closer to product management than pure research. Excellent read!

🔹 IBM drops 299-page playbook on creating business value with AI
→ AI Value Creators is a free blueprint from IBM execs on moving beyond prototypes to real adoption. Covers business impact, governance, operating models, and cultural roadblocks.

🔹 Anthropic warns Claude misused in cybercrime ops
→ Important signal: advanced models are already being misused for fraud, phishing, and malware creation. Regulatory + defensive measures will become a core part of the agent ecosystem.

🔹 AWS releases 80+ page guide on building AI agents
→ Reads like AWS’s playbook for cloud-native agent systems. Covers frameworks (Strands, LangGraph, CrewAI, Bedrock), protocols (MCP, A2A), security foundations, and scaling guidance. Centered on AWS stack, but an excellent resource for leaders and builders going hands-on.

🔹 Booz Allen releases framework for enterprise GenAI apps
→ Outlines a six-layer tech stack (infra → platform → LLM → data → agents → UI) with guidance on deployment models, LLM orchestration, and real-time data pipelines. Adds LLMOps practices like drift monitoring and retraining, plus GRC frameworks and human-in-the-loop guardrails.

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

There hasn’t been much hard evidence that AI is truly impacting jobs — the debate has been louder than the data.

In recent months, one camp has argued that AI is already eliminating opportunities for new graduates. Derek Thompson suggested that the slowdown in job-finding among college grads might be the first sign of a “job-pocalypse,” a narrative quickly picked up in the press. Others pushed back, pointing to alternative data sources. Research from Sarah Eckhardt and Nathan Goldschlag of the Economic Innovation Group, for instance, found no detectable effect of AI on recent employment trends.

Against this backdrop, a new Stanford study offers some of the clearest evidence yet that entry-level jobs are starting to shift. It’s the first study to use millions of payroll records to track AI’s real impact on work. The researchers find that since late 2022, employment for 22–25 year olds in AI-exposed jobs like software development has declined by around 13 percent. More experienced colleagues in the same roles remain unaffected.

Interestingly, the same trend appears in other roles that are heavily exposed to AI — such as customer service representatives (20%), which are widely considered especially vulnerable.

The natural next question is: What about jobs that aren’t as exposed to AI? Take for example home health aides — a role that requires physical presence and direct, hands-on work with patients. In these fields, the trend flips. Unlike software or customer service, it’s the youngest workers who are seeing the fastest job growth.

The study finds: For 22–25-year-olds, employment is rising in the least AI-exposed jobs (like health care) but declining in the most AI-exposed jobs (like software development or customer service).

In contrast, for older age groups, there is no meaningful divergence in employment patterns by AI exposure.

The decline is concentrated in jobs where AI automates rather than augments work.

The six core findings of the study

→ Entry-level workers in AI-exposed jobs are experiencing significant employment declines.
→ Experienced workers in the same roles remain stable or growing.
→ Automation-driven occupations (coding, accounting, customer support) are hit hardest, while augmentation-driven roles continue to expand.
→ Firms adjust through hiring freezes, not wage cuts.
→ The pattern holds across industries, not just tech. The timing overlaps with the breakout of generative AI in late 2022.

My perspective

I’m cautious about blaming this entirely on AI. High interest rates, cooling trade, and the unwinding of failed tech bets also weigh on hiring. The study controls for some shocks, but disentangling causes is difficult.

Still, the pattern is clear: fewer junior roles, tougher entry into AI-exposed fields.

And disruption won’t stop at 22–25-year-olds. Mid-career workers in their 30s and 40s may look safe now only because firms freeze junior hiring instead of firing experienced staff. Pressure is masked — until AI tools mature or restructuring cuts deeper.

The Implications

If junior positions erode, organizations lose pipelines for training and mentorship. Tacit knowledge, which AI cannot replicate, risks being stranded at the senior level.

• Career entry points shift upward — harder for young professionals to start.

• Training and applied learning become even more critical.

• Inequality grows if networks replace open entry paths.

• Industries risk stagnation without new talent pipelines.

The bottom line: the first rung of the ladder isn’t gone, but it’s higher.

The Opportunity

This isn’t “AI takes all jobs.” It’s simpler:

• AI compresses the entry level when used for automation.

• AI expands the pie when used for augmentation.

Which path we take is still a choice.

What To Do (tactical)

Don’t get comfortable. Today’s pressure shows up at the entry level, but it will climb. Freezes in junior hiring eventually reshape the whole career ladder. Invest in augmentation skills — workflows where AI adds leverage (automation orchestration, oversight, decision-making), not where it replaces.

Position yourself in roles where judgment, strategy, and leadership matter most. These are harder to automate and more valuable to scale. And keep in mind that fundamentals still compound:

• Writing — every AI interaction starts with language.

• Math / data intuition — know how data scales, transforms, and arranges.

• Problem-framing — the core skill AI still can’t replicate.

And if you’re in a leadership position, it’s time to rethink talent pipelines. The organizations that create new on-ramps — through mentorship, apprenticeships, and applied learning — will be the ones that attract and nurture the next wave of AI talent.

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

Bright Data just shipped an MCP server that finally makes large-scale web access for AI agents practical — and it’s FREE to start.
→ Data access is still the #1 bottleneck for agent workflows
→ Scraping infra is expensive, fragile, and gets blocked
→ And without real-time grounding, LLMs hallucinate

How it works:
Runs as an MCP server compatible with Claude and any MCP-ready LLM. Which gives your agents:
→ Real-time search, navigation, and extraction without getting blocked
→ Geo-unlocking + browser automation for dynamic sites
→ Structured, compliant datasets (text + multimodal)
→ Plug-and-play integration into RAG pipelines, agents, and LLM workflows

And the best part: you get 5,000 free monthly requests — enough to prototype serious agent workflows without building your own scraping stack.

Under the hood, it’s powered by Bright Data’s AI-ready products (Search API, Unlocker, Agent Browser, Web Archive API).

→ Try it: AI Access Toolkit
→ MCP server: Install here