This is what Junjo is for https://t.co/UiuVIMOWnk

2 more GitHub stars LFG!!!

Docs update Easy concurrency in Junjo workflows https://t.co/rP9vwbIJiK

Your prompts are not your moat. Your evals are. Your evals inform the prompt. Models change. Prompt maintenance is ongoing. Evals are forever.

I started making Junjo because existing workflow frameworks were too difficult to build and test complex workflows with. Converting my voice-controlled AI workout tracker to Junjo now that the API feels really stable. First node completed, complete with dozens of evals and

got a blue check for my library so you could say things are getting serious

Eval-driven development doesn't need special tools. It should happen inside your codebase. With junjo + pytest + copilot, it's easy to generate new synthetic tests for every node in your AI graph workflow. Change prompts and models, and re-execute every test in parallel FAST!

Cloudflare pages is one of the best deploy experiences.

I finally got the initial version of my Python AI Graph Workflow library up on PyPI and pushed up a sphinx-docs site for it: https://t.co/hOI0lvPfyr

> pip install junjo Junjo is officially released as an open source package. Still working on docs site and intro video, but the ai_chat example is complete. (@junjo_ai) https://t.co/uReeTcNZo6 https://t.co/zNZ38bXOhW

Exception visualization makes debugging why your workflows are failing much easier.

Why LLM workflow execution frameworks are becoming so important. Concurrently executing a dozen specialized 8b model requests is approximately 100x faster and cheaper than trying to have a 300b thinking model accomplish the same task.

Added a new view to the Junjo AI Workflow state machine traces / debugger - a flat chronological list of state updates. It will still visualize the workflow node that initiated the state updates.

The more I use LLMs for coding, the less code I actually have them write, and the more I just ask it about ways to do things.

Noticing myself adopting a certain rhythm in AI-assisted coding (i.e. code I actually and professionally care about, contrast to vibe code). 1. Stuff everything relevant into context (this can take a while in big projects. If the project is small enough just stuff everything

Today is the start of a new era of natively multimodal AI innovation. Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality. Llama 4 Scout • 17B-active-parameter model

. @xai when will the Grok3 API be available? Any turbo API models? There are many use cases for Grok over others.

Progress on visualizing telemetry traces and graph execution structure for @junjo_ai workflows. A simple new pattern for concurrently executing a list of nodes. Left: Junjo Server UX Right: ReactFlow and Mermaid diagram rendering.

eval-driven-development with pytest and @junjo_ai LLM assert example. Iterate on prompts with hundreds of tests for every AI workflow node. Test for 100% accuracy all the way through your logic. Next: automate test cases harvesting from production telemetry.

Constructing more test app examples for Junjo. It's a beautiful thing when the most basic auto-complete LLMs recognize the context of graph / edge changes for auto-compete.