Why
We’re building the first foundational version of agentOS to understand what’s possible in autonomous software-engineering workflows and gather real feedback from early users. This version is meant to give us an initial “feel” for the product, the UX, and the agent capabilities.
What
Create an MVP of agentOS focused on software engineering, where the system acts as a highly capable “project-lead coworker.” It should be able to:
- Interact with the user conversationally
- Take a task and independently execute work
- Produce meaningful output across design, architecture, code, infra, and deployment
Success
A fully working pipeline from:
- User asks for a “build roadmap tool”
- → agentOS designs it
- → architect + generates code with a clean structure
- → makes sound infrastructure decisions
- → deploys a functioning version
If this pipeline reliably produces a decent result end-to-end, Phase 1 is successful.
Scope
Deliver a runnable web application where users can:
- Create a new workspace (a.k.a. “company”)
- Start the first task: build an MVP of a simple roadmapping tool (reference spec)
- Trigger and monitor an agentOS execution
Core requirements:
- Integrations: Codex, Claude Code, main orchestration LLM
- Authentication: Vercel, GitHub, or Convex (pick one for MVP)
- Tech Stack: Highly opinionated, strict constraints; provide a starter repo so the agent has a defined execution environment
- Observability: Show real-time agent steps, decisions, errors, and progress
- Feedback Loop: Agent pings user asynchronously for clarifications / decisions
- Safety Controls: Hard token-spend limit per workspace to prevent runaway executions