Deepagents Releases Open Source Claude Agent SDK
Today's top 14 insights for PM Builders, ranked by relevance from X, Blogs, and YouTube.
Deepagents Releases Open Source Claude Agent SDK
#1 š
Harrison Chase launched Deepagentsāan open-source Claude Agent SDK available at github.com/langchain-ai/deepagents for building and running Claude-powered agents.
#2 š
Garry Tan shows how to install OpenClaw or Hermes from his gbrain repo and have it running on WebRTC or your Twilio number in under 30 minutes. He likens it to a Homebrew computer club for personal AI.
#3 š Simon Willison
MLX audio: uv run recipe for transcribing an audio file (Gemma 4 E2B) - A reproducible uv run command for transcribing audio on macOS using the Gemma 4 E2B model with MLX and mlx-vlm, plus an example transcription of a 14-second .wav file showing some misrecognitions. The post includes links to the model, tool, and demo audio file.
#4 š Simon Willison
QuickJS Python Sandbox ā Investigation Report (Research) - A research note exploring the quickjs Python package that implements an asyncio-compatible JavaScript sandbox with resource controls and the ability to expose synchronous and asynchronous Python functions (including async httpx fetches) to JavaScript.
#5 š
Guillermo Rauch suggests creating an X group chat with your most demanding customers and engineering leads to gather direct feedback and ship rapid @v0 updates. This tight loop forces you to face critiques head-on and iterate products quickly.
#6 š
Garry Tan highlights that even though OpenClaw wonāt auto-generate tests while building your software, you should explicitly prompt it to write tests for you.
#7 š
Alexandr Wang highlights Muse Sparkās strength in sourcing and analyzing open-source data, demonstrated by his century-long analysis of global GDP share.
#8 š
Alexandr Wang spotlights Contemplating mode, Meta AIās new feature that delivers enhanced deep reasoning for tackling the most complex queries.
#9 š
Thariq notes that Claude Code now includes a TurboTax connector, enabling direct taxāprep automation within the AI platformāand heās relieved his procrastination paid off!
#10 ā¶ļø
Figma CEO on How to Get Good at Design in the AI Era | Dylan Field
Peter Yang
Dylan Field outlines Figmaās AI-driven design workflow, including Figma Makeās AI agent-generated divergent canvas iterations, direct-manipulation superiority over prompting, and the Figma MCP plugin for seamless code-to-design roundtrips.
- Gemini 3.0 and 3.1, when prompted with complex instructions and reference images, deliver high-fidelity visual design outputs on the Figma canvas.
- Figma released approximately 200 features in the previous year and plans to deliver an even greater magnitude of user-impactful features and larger initiatives in the current year.
- 60% of design files in Figma are created by non-designers through Figma Make and the platformās open canvas for rapid divergence and convergence loops.
#11 š
Boris Cherny says Claudeās default dialogue style was switched to Medium after users complained about high token usage. He clarifies that response lengths havenāt been cut and future tweaks will follow user feedback.
#12 š
Harrison Chase introduced LangChains create-agent as a super-minimal agent SDK, contrasted with DeepAgents as a batteries-included alternative. He also highlighted how middleware can extend and customize both frameworks for advanced workflows.
#13 š
Peter Yang interviews Figma CEO Dylan Field on AIās role in design, the creativity-versus-structure debate of design systems, and Figmaās pivot as two-thirds of its users become non-designers.
#14 š
Guillermo Rauch notes that poke.com cleared the AGI benchmark threshold last year and has been auto-checking him into every domestic flight for over six months.