BX Forge

BX Forge Inspiration The challenge was "build an agent that builds agents" — and we noticed almost every entry would stop at generating a skeleton. That's the easy 80%. The real opportunity was the go-further: don't just hand someone files, prove they work. We pushed that idea as far as it would go — from "generate one agent" to "connect your actual GitHub repo, generate a crew tailored to it, and have that crew open a real pull request."

What it does Describe a problem in plain English and BX Forge designs an agent spec, scaffolds a real runnable BoxLang AI project, runs it, and has a second AI act as judge — pass/fail, score, reasoning. Connect a GitHub repo and it goes further: BX Forge inspects your codebase and generates a repo-tailored multi-agent crew (planner, implementer, reviewer, tester) wired to real repo tools, which runs for real and opens an actual pull request. A standalone CLI does the same generation with no web app — the reusable dev tool the challenge's community goal asks for.

How we built it BoxLang backend on the MiniServer, BoxLang AI (aiAgent/aiTool) for real model calls, Vite + React + Tailwind frontend. Pipeline: SpecAgent (description → structured spec) → Scaffolder (spec → real files) → AgentRunner (executes the generated project, clones repos for the crew path) → Evaluator (AI judge). Four of us split ownership by layer, with GitHub Issues and a written contracts doc as the single source of truth so parallel work didn't collide.

Challenges we ran into The real BoxLang AI syntax wasn't documented anywhere online — we had to reverse-engineer it from the workshop deck after early templates guessed wrong and no generated agent could run. The bx-ai module itself was never installed for a stretch of the event, so the entire AI path silently did nothing until we debugged it from first principles. BoxLang's CLI doesn't respect shell quoting, corrupting multi-word input with zero error. And GitHub's OAuth redirect kept wiping our app's state, bouncing users back to square one after a successful login.

Accomplishments that we're proud of The full loop is real, not simulated — live model calls, a generated project that actually executes, an AI judge verdict with real reasoning, and a real branch, commit, and pull request opened against a real repo, verified live. We kept the original single-agent flow fully working as a safety net through the entire pivot to the repo-crew feature.

What we learned Undocumented APIs need empirical verification, not assumptions. A tight, explicit ticket-and-contract system let four people build against each other's in-progress work without merge chaos.

What's next for BX Forge Finish the code-review/vulnerability-scan clarification flow, expand repo-tool coverage, add real model selection, and publish the CLI as an installable BoxLang module.

Built With

• boxlang
• javascript
• react