Five Repos, One Warning About the Ontology Crowd

• langgenius/dify (GitHub) - 144,296 stars. This is a production-ready platform for building agentic workflows without writing the orchestration layer yourself, and it's the open source answer to the proprietary agent builders. That star count says a lot of teams have already made the swap. Written in TypeScript, so it plugs into the stack most of you already run. Evaluating agent infrastructure? Close this gap first.

• alibaba/open-code-review (GitHub) - Alibaba open sourced the code review tool it runs internally, and for anyone shipping software daily, this is the most useful thing on the list. It mixes deterministic pipelines with LLM agents, supports both OpenAI and Anthropic models, and comes with rulesets for the security issues that actually show up in real codebases. Battle-tested at Alibaba's scale is a claim worth taking seriously, because most code review tools are battle-tested at nobody's scale. Wire this into your CI and you get a second reviewer that never gets tired on a Friday afternoon.

• obra/superpowers (GitHub) - 219,579 stars and still trending, with over a thousand new stars in a single day. It bills itself as an agentic skills framework and a development methodology, written in Shell. The traction is real. What's thin is the detail on how it actually integrates with the tools most of you are running, so go in expecting to do some reading before you know if it fits your workflow.

• earthtojake/text-to-cad (GitHub) - Nearly 10,000 stars for agent skills built specifically for CAD, robotics and hardware design. This is the interesting kind of agentic tooling, the kind that goes deep on one domain instead of promising to do everything. Most agent frameworks stop at writing code and answering emails, but this one is aimed at people building physical things. A much smaller, much more useful lane.

• huggingface/speech-to-speech (GitHub) - Hugging Face's repo for building local voice agents on open source models picked up 627 new stars in a single day, which tells you where developer attention is heading next. Voice is the interface everyone keeps predicting and nobody quite ships. This gets you local, which matters if you care about latency, or if you don't want your users' voices going to a third-party API. Early days. Clone it and run it yourself before you decide it's not for you.

The Latent Space piece on ontologies making a comeback is worth sitting with, because it names something I keep seeing without a name for it. Agents are probabilistic. Businesses are not. You cannot ship a support workflow that's right ninety per cent of the time and shrug at the other ten, and better models won't fix that. What fixes it is a harder boundary around what the model's allowed to decide.

Same lesson, buried in Alibaba's code review tool this week. It doesn't let the LLM freewheel over a diff. Deterministic rules run first, and only the ambiguous cases go to the agent. Match the model to the job. Don't ask frontier reasoning to do what a rule could do for free.

The ontology crowd spent a decade being told the semantic web was dead, and it turns out it was just early. I've heard that one before.

Have you built a hard boundary around an agent this week, or are you still trusting the prompt?

Tell me in the comments.

Forward this to one person who should be using AI better than they are. Reply with what you built, tried or broke this week. I read every one.

Gareth, founder of The Anthropic Stack (theanthropicstack.com)

P.S. Shipping something this week? Railway gives Anthropic Stack readers $20 in credits to start: theanthropicstack.com/railway. It is a referral link, so I earn a small cut if you stay. I only point you at tools I run myself.