How one person built 18+ projects in 4 days with a multi-agent Claude Code architecture.
Architecture
The Process
Spec in Notion Document the idea, architecture, acceptance criteria. Claude Code reads the spec via MCP tool, so it knows the goal before writing a line.
Create a Coder workspace Fresh remote machine with the repo cloned. Each project gets its own isolated environment so agents never step on each other.
Start Claude Code Give it the Notion spec + CLAUDE.md context. The combination of project instructions and a clear goal means it can work autonomously for hours.
Let it build Claude Code creates files, runs tests, iterates on failures, installs dependencies, and self-corrects. You watch the logs or go start the next one.
Communicate via file bus The Bridge session coordinates agents by writing/reading files in /tmp/claude-bus/. Status updates, requests, completed signals all flow through plain text files.
Pull work to main branch Bridge pulls completed code from workspace branches, resolves any conflicts, and prepares the deployment config.
Deploy via cake.sh Build the Docker image, push to ECR, update the ECS service. One command: ./cake.sh deploy minime
Iterate Kirk reviews, gives notes, Claude adjusts. Rapid feedback loops. The human does product thinking; the agents do the building.
Tips for Solo Use
No multi-machine setup needed. One Coder workspace is enough.
Use Claude Code's built-in Agent tool for parallel work within a single session
Use isolation: "worktree" for agents that modify files - prevents conflicts
Use /loop for recurring status checks on long-running tasks
Keep your main conversation for coordination, delegate research and building to sub-agents
Never block the primary thread. If a task takes more than 10 seconds, send it to a background agent. You should always be able to talk to Claude while work happens. Use run_in_background: true on agents so you get notified when they finish instead of waiting.
Break work into discrete tasks - one agent per task, clear inputs, clear outputs
Let agents commit to branches, you merge - treat them like junior devs on PRs
CLAUDE.md is your force multiplier - the better your docs, the less you repeat yourself
Voice-to-text works. Kirk dictated most instructions. Claude interprets intent, not exact words. Speak naturally - corrections are cheap.
What Worked / What Didn't
Worked
Parallel agents on separate machines for independent projects