← Back to archive

Wednesday, May 20, 2026

Anthropic acquires Stainless (+ 30-agent chaos)

Anthropic just acquired Stainless to supercharge Claude's agent connectivity through better SDKs (smart play for the API wars), while Qwen dropped their 3.7 preview with open-source multimodal MoE models that are apparently crushing benchmarks. Meanwhile, someone built "Gas Town"—a system running 30 concurrent AI agents for industrial-scale software development, which is either the future or absolute chaos (probably both). Would you trust 30 agents to build your production code?

Top Stories

1
Anthropic Acquires SDK Startup Stainless

Anthropic

Anthropic acquired Stainless, the SDK generation platform behind all official Claude SDKs, to strengthen agent connectivity as AI shifts from answering questions to taking action. The move reinforces Anthropic's Model Context Protocol (MCP) strategy for enabling AI agents to connect with external tools and data.

anthropicagentsmcpsdk
2
Welcome to Gas Town

Medium

Gas Town is a Kubernetes-like orchestrator for managing 10-30+ concurrent AI coding agents, featuring persistent Git-backed workflows (MEOW stack) that enable autonomous multi-agent software development. The system is production-ready but expensive and requires advanced AI-coding expertise.

agentsorchestrationclaudeworkflow-automation
3
Qwen3.7 Preview lands on Arena

Thread Reader App

Alibaba's Qwen releases a comprehensive suite of open-source models including Qwen3.5-397B (multimodal MoE with 19x faster decoding), Qwen3-Omni (first end-to-end omni-modal AI), and enhanced image generation/editing models, all Apache 2.0 licensed and optimized for agent applications.

qwenopen-sourcemultimodalmoe
4
Fine-Tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for Robot Video Generation

Hugging Face

NVIDIA's guide demonstrates how to fine-tune the Cosmos Predict 2.5 world model using LoRA/DoRA adapters to generate synthetic robot training data efficiently on single GPUs, achieving significant improvements in video quality and task adherence with minimal compute compared to full fine-tuning of the 2B-parameter model.

roboticsworld-modelsnvidiafine-tuning
5
The Unreasonable Effectiveness of HTML

HTML-based AI agent outputs enable spatial layouts, interactive prototypes, inline diagrams, and custom editors that communicate far more effectively than linear markdown, creating tighter iteration loops and better human-AI collaboration.

agentsai-interfacesdeveloper-toolsprompt-engineering

Keep Reading

Industry Voices

Enjoyed this issue?

Get daily AI intel delivered to your inbox. No fluff, just the stories that matter.