Parallel Coding Agents Need Reviewable Workspaces

·https://www.conductor.build/
articletoolaideveloper-toolsagent-loopsgit-worktreescode-reviewmacos

Conductor's worktree-per-agent model maps to joelclaw workload review boundaries for local agent swarm runs.

Conductor is a macOS app for running Claude Code, Codex, and Cursor agents in parallel. The clever part is it makes every agent stream a review object: a branch, a Git worktree, a terminal, a running app, a chat, a diff, and a pull request path.

Joel’s Slack note paired Conductor with “Claude Code swarmtools.” The concrete pattern worth remembering is not “more agents.” It’s separate lanes with visible merge gates. Parallel agent work gets useful when each lane can be reviewed, merged, archived, or thrown away without poisoning the main checkout.

That maps cleanly to joelclaw. The agent loop shape already wants bounded tasks, reviewer steps, and resumable child work; Conductor shows the local desktop version of that same discipline. Fan out experiments with Claude Code or Codex, keep each in its own Git worktree, and only promote the diff that earns it.

The risk is right in the Conductor FAQ: agents run with the same permissions as the user, so this is development isolation, not a security boundary. Still useful. Just don’t confuse “separate workspace” with “safe sandbox,” because that’s how the machine gets spicy in the bad way.

Key Ideas

  • Conductor treats each coding-agent task as an isolated workspace with its own branch, Git worktree, commands, chat, diff, and review flow.
  • Conductor’s parallel agent docs make a useful distinction: use multiple workspaces when work can land independently, and use multiple agents in one workspace when they need shared branch state.
  • Conductor’s Git worktree model is the boring infrastructure move that makes agent swarms less chaotic: separate files on disk, shared repository history, and one branch as the review unit.
  • Conductor supports Claude Code agent teams through CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, which lines up with Joel’s “swarmtools” note.
  • The model fits joelclaw best at the operator edge: local fan-out, visible diffs, explicit review, then hand the winning branch back to the durable system.
  • Workspace isolation is not sandboxing; Conductor’s FAQ says agents run with the user’s permissions.

Links

← All discoveries