03/2026

AI Agents

Max-agents | Coding

Agent system for Claude Code to deliver software

What is it about?

Agent system for delivering projects and products from researching an idea to launching it for audiences.

Dashboard above.

Five agents with controlled handovers between. Each achent orchestrates sub agents, tools and skils independently:

  • Researcher
  • Prototypes
  • Architect
  • Builder
  • Launcher

In addition, a Patcher for end-to-end small jobs.

What’s different compared to other systems is that sub-agents are specialised according cognitive needs, not sections or languages.

Best use with the Max plan of ClaudeCode.

Why I’m building it?

To have superpowers to build. Its working! Lot of iteration still needed to get ir better.

How am I building it?

I had earlier a 15 agent manual handover system for coding. I was the bottleneck. I asked CC to compare my system with ‘Everything Claude Code” and interview what’s important in both agents. In that process this version of coding agents was born.

What’s my goal?

I’m proving it works with two quite ambititius projects:

  • Foresighter
  • PITs

Tech Corner

Languages

  • Bash (orchestration, all 18 scripts)
  • Markdown (agent instructions — 107 files)
  • Python (dashboard server, metrics)
  • HTML/CSS/Vanilla JS (web dashboard)

AI & LLM

  • Claude Code CLI
  • claude-opus-4-6: Architect, Builder orchestrator, Patcher
  • claude-sonnet-4-6: Builder workers, reviewers, Launcher
  • claude-haiku-4-5: Lightweight/mechanical tasks
  • Agent tool for parallel sub-agent dispatch (up to 20+ simultaneous)
  • Git worktree isolation per parallel task

Agents

  • 6 main agents: Researcher, Prototyper, Architect, Builder, Launcher, Patcher
  • 50+ sub-agents across all roles (builders, reviewers, testers, utilities)

Integrations & tools

  • GitHub (PR creation, releases)
  • Vercel (preview + production deploy)
  • Supabase (DB migrations)
  • Git (branch management, worktrees)
  • Claude Code CLI (--append-system-prompt)

Infrastructure

  • Handoff system: JSON files between agents
  • Audit log: append-only event trail
  • Checkpoint/restore for long runs
  • Task graph with file ownership conflict detection
  • Dashboard: terminal (ncurses) + web (Python HTTP server)

Lines of code

  • Bash scripts: ~5,500
  • Agent .md files: ~12,100
  • Dashboard: ~1,300
  • Total: ~21,000
← Back to projects