Back to journal
ai MAY 19 · 2026

Conductor: How I Run a Whole Team of Claude Agents From One Mac

The bottleneck moved. A single Claude Code agent is already faster than me. The limit is how many I can supervise without losing the plot. For a long time that was one.

With Conductor, it’s closer to ten.

What it is

A Mac app that runs Claude Code agents in parallel. Not a new model, not a new IDE — a harness. It reuses my existing ~/.claude setup verbatim: skills, hooks, MCP servers, RTK, CLAUDE.md. No re-onboarding.

One overview, many agents, many projects

Every workspace across every project, one window, status at a glance.

Conductor workspaces overview

Right now: one agent writing this post, one on a Jekyll deploy script, one on a flaky test in another repo, one drafting a PR. Three projects. I know who’s blocked on me without clicking in.

The loop: plan → implement → review → PR → merge → archive

Every task moves through the same six steps. My job is two of them — approving the plan, reviewing the diff. The rest is the agent.

Worktrees + setup/run scripts

Every task gets its own git worktree. Own branch, own checkout. Three versions of the same site can run side by side with no collisions. Each workspace runs my project’s setup script (deps, env) and run script (dev server) on demand. Testing a branch is one click.

Staying out of the IDE

The unexpected win. When the editor is in front of me I twitch — rename things, fix a semicolon, lose an hour. Conductor keeps me one step back, on intent and review.

When I actually need code, the diff viewer is right there. “Open in IDE” and “open terminal in this worktree” are one click each. No access lost — just no default pull.

GitHub issues in, PRs out

Point Conductor at an issue. It opens a workspace with that context, plans, executes, opens a linked PR. Backlog turns into throughput.

So is it smarter than Claude?

No. That’s the point. Conductor isn’t a better model — it’s the safety rails (worktrees, PRs, structured review, archive) that let me run a lot of Claude at once without it turning into chaos.

If your bottleneck moved from “the AI is too slow” to “I can only supervise one of these at a time,” try it.

Follow Brian Porter on LinkedIn →
← Previous
Brian Porter

Written by

poornerd

CTO at an automotive data company in Munich. Co-founder of SiteForce AG. Four decades writing software and shipping production systems.

Keep reading

Related essays