AI Company — Autonomous Software Company

From idea to a real GitLab repository with production code — in under 3 minutes, powered by Gemini 2.5 Flash

Comment

Inspiration

Building a software company requires months of planning, dozens of meetings, and hundreds of hours of coordination between product managers, architects, developers, QA engineers, DevOps engineers, and marketing teams. We asked: what if a single AI could replace all of that? What if you could type an idea and get a fully planned, estimated, and executed software project in minutes?

What it does

AI Company is an autonomous software company powered by Gemini 2.5 Flash. You give it any project idea and six specialist AI agents go to work simultaneously:

  • PM Agent creates user stories, sprint plans, and feature priorities
  • Architect Agent designs the system architecture and database schema
  • Dev Agent estimates hours, recommends tech stack, and generates production code
  • QA Agent creates test plans and identifies security risks
  • DevOps Agent plans infrastructure, CI/CD pipelines, and cost estimates
  • Marketing Agent builds go-to-market strategy, KPIs, and launch plan

After planning, the system:

  • Generates 5 real production code files (server.js, schema.sql, App.js, docker-compose.yml, README.md)
  • Creates a real GitLab repository with code files, issues, and Sprint 1 milestone automatically
  • Runs an internal team meeting where agents debate trade-offs in real time
  • Optimizes the budget with a slider (10-70% reduction) and AI renegotiation
  • Plans workforce changes with Brooks Law analysis
  • Saves every project to Firestore for agent memory across sessions
  • Logs every API call to Arize for AI observability

How we built it

  • Gemini 2.5 Flash via Google Gen AI SDK as the AI brain
  • FastAPI Python backend with 6 specialized endpoints
  • GitLab REST API for repo creation, file uploads, issue creation, and milestone management
  • Google Firestore for persistent agent memory across sessions
  • Arize AI for observability — logging latency, prompts, and responses for every Gemini call
  • Google Cloud Run for deployment

  • Vanilla HTML/CSS/JS frontend with animated agent cards

    Challenges we ran into

  • Orchestrating 6 agents with shared context without using Agent Builder UI

  • MongoDB SSL compatibility issues led us to switch to Google Firestore

  • Managing Gemini API rate limits with retry logic

  • Getting GitLab repo creation, file push, and issue creation to work reliably in one flow

    Accomplishments that we're proud of

  • Built a fully working autonomous software company from scratch

  • Real GitLab integration that creates repos, pushes code, creates issues, and sprint milestones

  • Budget optimizer that uses Gemini to intelligently renegotiate scope and staffing

  • Brooks Law workforce planner that reasons about adding developers

  • Agent memory with Firestore that remembers all past projects

    What we learned

  • Multi-agent orchestration is more powerful with clear system prompts than complex routing logic

  • Gemini 2.5 Flash is surprisingly good at playing multiple specialist roles simultaneously

  • Real tool integrations (GitLab) make demos 10x more impressive than simulated ones

  • Brooks Law is real — the workforce planner correctly warns when adding too many developers slows things down

    What's next for AI Company — Autonomous Software Company

  • Add MongoDB integration for richer agent memory

  • Integrate Elastic for searching past projects

  • Add Dynatrace for infrastructure monitoring recommendations

  • Fivetran data pipeline setup via Marketing agent

  • Multi-user collaboration — multiple people interacting with the same AI company

Built With

  • arize-ai
  • fastapi
  • gemini-2.5-flash
  • gitlab-api
  • google-cloud-run
  • google-firestore
  • google-gen-ai-sdk
  • python
Share this project:

Updates