Agent Swarm Experiments
Three experiments from 2025 in running Claude Code autonomously, in Bash and Node.
In date order, they move from single-agent memory, to orchestrating many agents, and back to the memory-first approach that became Dev Framework. They are part of the personal practice behind bringing AI tooling to our development teams.
- Autonomous agents
- Orchestration
- Agent memory
Memory first
ClaudeCodeAutonomousAgentSetup came first, and it is not an orchestrator. It is a set of install scripts that scaffold one agent's session memory: CURRENT-STATUS.md, NEXT-TASKS.md, decision logs, and per-session notes, plus CLAUDE.md instructions that tell the agent to keep all of it current.
Many agents
AutoASS turned a Claude MAX subscription into five specialized agents, Planner, Frontend, Backend, Integration, and QA, building applications autonomously on Windows, WSL, and Linux.
Then came swarm: a Node CLI orchestrator for multiple Claude Code agents collaborating on development tasks, with agent templates and git worktrees.
Back to memory
The arc ends where it began. The memory-first approach from the first experiment is the one that carried forward, and it became Dev Framework.