📡 The Agent Economy Arrives

DeepMind Genie: When AI Learns to Control Worlds

The "Sora moment" proved AI could dream consistent worlds. The next frontier is proving AI can control them.

DeepMind's Genie represents a paradigm shift in world models—transforming static video into interactive, playable environments. Unlike passive video generation (beautiful to watch, impossible to steer), actionable world models understand agency: not just that a cup fell, but what caused it to fall.

Why it matters: For AI to truly understand reality, it must understand causality and control. This is the essential building block for physical AI, robotics, and simulation.

The Coding Agent Price War

The AI coding revolution has a dirty secret: it's expensive. Claude Code charges $20-$200/month with strict rate limits. But a rebellion is brewing.

Goose, Block's (Jack Dorsey's company) open-source AI agent, offers nearly identical functionality to Claude Code—for free. It runs entirely locally, works offline, and has no rate limits.

• 26,100+ GitHub stars
• 362 contributors
• Works with any LLM (Anthropic, OpenAI, Google, or local models via Ollama)

The implications are profound: the $200/month era for AI coding tools may be ending. Open-source models like Qwen2.5 and Kimi K2 now benchmark near Claude Sonnet 4 levels—and they're free.

Railway: The Cloud Built for Agents

Railway just raised $100M to challenge AWS with infrastructure designed for AI-native development.

Their pitch: legacy cloud was built for humans. AI coding agents generate code in seconds—waiting 2-3 minutes for Terraform deployments is unacceptable.

31% of Fortune 500 now use Railway. The founder's vision: "The amount of software that's going to come online over the next five years is unfathomable—we're talking a thousand times more software. All of that has to run somewhere."

Microsoft CORPGEN: Digital Employees Arrive

Microsoft Research released CORPGEN, a framework for autonomous "digital employees" that can manage dozens of concurrent tasks—mimicking real knowledge workers.

Key innovations:

• Hierarchical planning: Breaks objectives into daily goals → moment-to-moment decisions
• Tiered memory: Selective recall instead of retaining everything
• Experiential learning: Agents store completed tasks and reuse patterns for similar work
• 3.5x higher completion rates than baseline agents under multi-task loads

The striking finding: Memory and retrieval, not raw model capability, may be the key bottleneck for real-world agents.

Listen Labs: AI Replaces Market Research

Listen Labs raised $69M to scale AI-powered customer interviews.

Their platform:

• Recruits from 30 million person network
• AI conducts in-depth interviews with follow-up questions
• Delivers executive-ready reports in hours, not weeks
• Eliminates survey fraud through identity verification

Microsoft, Sweetgreen, and Chubbies are already customers. Traditional market research takes 4-6 weeks; Listen delivers insights in days.

NVIDIA: Game Inference + Qwen3.5 VLM

NVIDIA announced new tools for:

• In-game inference: Running AI models alongside game engines efficiently
• Qwen3.5 VLM: Alibaba's ~400B parameter native multimodal agent with MoE + Gated Delta Networks
• Telco reasoning models: Autonomous network management via NVIDIA NeMo

Quick Hits

The Week Ahead

• World Models: Expect more Genie-style interactive AI research
• Agent Economics: Watch for more "AI-native" infrastructure plays
• Open Source vs Proprietary: The gap continues to close