Announcing Sgai, or How We Stopped Writing Code (Mostly)

Sgai Is Pronounced Sky

Ulderico built Doc Holiday’s release notes writer, and then asked Sgai to find and fix its bugs...while he worked on something else.

“It was like I had a superpower,” he said.  “I was multithreaded.  It was AMAZING!”

Ulderico, like all of us, has used all the tools.  We had GitHub Copilot, Cursor, and Claude Code licenses simultaneously.  They’re all great.  But we kept asking:

• “Well what if I just got the LLM to do that?”
• “What if I just had it ask another agent?”
• “What if I could run five of them at a time?”

Like Dan Shapiro outlined in his most excellent post, we started with spicy autocomplete and wandered through the dark factory.  Where we ended up, though, was a step beyond and into Open Fab / Makerspace territory.  

‍

What We Built

Sgai is an AI software factory built around a text box.  Type what you want. That's it.

"Build me a TODO app."

"Add export to CSV functionality."

"Fix the login bug where passwords with special characters fail."

Sgai's agents analyze the request and create a plan. Then they come back with questions.

"Should this be a single-page app or multi-page?"

"Should the CSV export include archived items?"

"Do you want user authentication or is this internal-only?"

Answer and the factory revises.  Approve and it builds.  While it works, a dashboard shows exactly what's happening. 

‍

Many Choices, All Excellent

There are many AI factories out there and all are great.  Like all software :) it’s a question of the best approach for the job: 

Agent Swarms ← Just throw tokens at the problem

• https://devin.ai/
• https://factory.ai/
• https://www.8090.ai/

Orchestrators ← Human is almost absolutely removed from the mix

→ Dark Factories (Specs-driven, lights-out; ideal when you know exactly what you want)

• https://github.com/strongdm/attractor

→ Complex Orchestrators (Fine-grained control, sophisticated workflows; ideal for teams requiring that level of control)

• https://github.com/steveyegge/gastown
• https://github.com/ruvnet/claude-flow

→ Interactive Factories (Conversational development; ideal when you discover what you want by seeing it)

• https://github.com/sandgardenhq/sgai (additional thoughts here)

Tool Help Drive Multiple AI Agents ← My job in the new age is to manage agents

• https://www.vibekanban.com/
• https://www.conductor.build/
• https://openai.com/index/introducing-the-codex-app/

Extend and TurboCharge the tool <- Human is very much present in the mix

• https://github.com/obra/superpowers - extends Claude Code
• https://github.com/code-yeongyu/oh-my-opencode -extends OpenCode
• Ralph Wiggums Technique (Geoffrey Huntley) - can extend any of Claude Code, OpenCode, Codex, Cursor, Amp

For more information about AI Factories and an analysis of each, please see this blog post.

‍

Check Sgai Out!

‍

Sgai releases today as a free, open-source project: github.com/sandgardenhq/sgai.  

Try Sgai: github.com/sandgardenhq/sgai
Documentation: github.com/sandgardenhq/sgai/tree/main/docs (generated by Doc Holiday, of course!) 
‍Questions? Ideas? Open an issue or start a discussion.