In January 2026, I started taking AI seriously. I bought some books on LLMs and current models to understand what I was using. I learned some prompting via Anthropic courses and some Udemy courses. Then I felt ready.
Salvador Is Born
I have a dedicated Mac mini for my agent. Salvador’s personality is very interesting; I’ve refined it over the months since April 2026, when I built him out. He is Claude-based and has full access to the Mac Mini. Claude Code and VS Code are the main ways I interface with Salvador.
I access him remotely in a few ways. On my phone or laptop, I use JumpDesktop, a secure, reliable remote desktop platform that enables me to seamlessly connect to any computer, anywhere in the world. I also configured Telegram to prompt Salvador and accept photos. Over time, I’ve trained him and created many, many skills that help me daily. I’ve got a dedicated agent for my job, Lews (C.S. Lewis, of course), who has access to and knowledge of everything at work.
Lately, I’ve been using swarms, as explained here, to accomplish more. It’s been a game-changer for getting more automation done faster and with higher quality.
Essentially, what I am calling a swarm (2024) is multiple Claude Code instances working in parallel, coordinating through a shared task list and messaging each other, with one session as the lead. It’s been handy for deep research and building software.
Configure
Getting started is pretty easy.
1. Enable it
Add the env flag to your settings.json (user: ~/.claude/settings.json, or project .claude/settings.json):
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}
Claude Code re-reads settings env on save, so you don’t strictly need to restart. Make sure you’re on a recent version (the simplified, no-setup flow is v2.1.178+).
2. Ask in plain English
Once enabled, in an interactive session you describe the task and the teammates you want, and the lead spawns them. Example:
Spawn three teammates to review PR #776:
- one on security
- one on performance
- one on test coverage
Have each report findings, then synthesize.
The lead creates a shared task list, spawns the teammates (each in their own context window), they claim/work on tasks, message each other, and the lead pulls it together.
3. Drive the team
- Specify count/models: “Spawn 4 teammates… use Sonnet for each.”
- Require plans: “Require plan approval before they make changes.”
- Talk to one directly (in-process mode): ↑/↓ in the agent panel to select a teammate, Enter to view/message it, Ctrl+T toggles the task list.
- Shut one down: “Ask the researcher teammate to shut down.”
- Reusable roles: reference a subagent definition — “Spawn a teammate using the
security-revieweragent type…”
Keep in mind
- ⚠️ Token cost: each teammate is a full Claude instance, significantly more tokens than one session. Best for research, review, debugging with competing hypotheses, and independent parallel modules and not sequential work or same-file edits.
- It changes normal delegation: while enabled, a subagent Claude names launches as a teammate (reports “idle,” not a returned result), so teams can form even when you didn’t ask. Set the flag back to
0to revert. - You’re in the VS Code extension → split-pane mode isn’t supported in the VS Code integrated terminal (nor Windows Terminal/Ghostty). The default in-process mode works fine here; split panes need tmux or iTerm2.
- Limits: one team per session, no nested teams, and
/resumewon’t restore in-process teammates.
Give it a try!
