Justy: Okay so episode five hundred. We're doing Conductor — it's a Mac app for running multiple coding agents at the same time.
Cody: Right.
Justy: Claude Code, Codex, Cursor — you spin them up in parallel, each in its own isolated workspace, and then you review the diffs before you merge. That's the shape of it.
Cody: So it's git worktrees under the hood, which is… fine, that's standard. The product is the orchestration layer and the review UI.
Justy: Yeah. And honestly I had a week where I was sick and working half-speed, so I'm going to blame that on why this took me a moment to land. But let me just catch you up on life first — you've been in DC, right?
Cody: Yeah, it's been humid as hell. I did a thing at work where we had to trace through a routing bug for eight hours straight and I just… I'm tired of routing bugs.
Justy: Fair. Anyway — so Conductor. The pitch is you're not running one agent and hoping it gets it right. You're running three agents on the same task, they each take a different path through your codebase, and you see which one works best.
Cody: Mm-hm.
Justy: But here's the thing that made me pause — if you're running Claude Code three times, you're just running Claude Code three times. There's no diversity there. The parallelism only helps if the agents are actually… different.
Cody: Exactly. Same model, same prompt, same token seed? You're just getting the same answer three times with different timing. The whole move only works if you're actually exploring the decision tree — different models, different reasoning paths, different prompts.
Justy: So the product assumption is that you'd spin up Claude Code on one task, Codex on another, Cursor on a third, and they'd each take a different approach.
Cody: Right. And that's reasonable — different models have different strengths, different biases about how to refactor or structure a fix. But the docs don't really emphasize that. They show you running the same agent type in parallel and reviewing the diffs.
Justy: Which is technically fine, but it's not making the case for why parallelism matters. It's just showing you the UI.
Cody: The actual value is the review interface. You're not trusting the agent output blindly — you see the diff, you see what changed, you decide whether to keep it. That's the human-in-the-loop part that actually matters.
Justy: Yeah. And I think that's where the product thinking is solid. Teams already using Claude Code are already spinning up agents and copy-pasting diffs into their codebase or reviewing them manually. Conductor just makes that workflow visible and codifies it. It's not a new religion about agents — it's plumbing for teams that are already doing this.
Cody: The payment model is clean too — you just use your existing Claude login or API key. No new billing layer, no auth rework. That's smart.
Justy: Right. And it's Mac-only right now, which is a real constraint for some teams, but that's a reasonable scope for a first product. You're not trying to be everything.
Cody: The worktree isolation is solid — each agent gets its own branch, its own chat history, its own terminal session. If one agent breaks something, it's isolated. You don't have to restart the whole repo state.
Justy: So what's the actual gap? Like, where does this not work?
Cody: Medium to large repos with complex dependencies. If you're running an agent on a five-thousand-file codebase with dynamic imports and generated code, the worktree is clean but the agent's understanding of what it can safely change is still bounded by context. Parallel agents don't fix that — they just let you see the failure modes faster.
Justy: Okay, so it's not magic.
Cody: Not magic. It's a better workflow for teams that are already comfortable with agents. If you're still deciding whether to use agents at all, this doesn't solve that question.
Justy: But if you're already using Claude Code in Cursor or you've got Codex running, this is like… it's the next thing you'd want. It's the orchestration layer you'd end up building anyway.
Cody: Yeah. Which is a solid product insight — solve the problem the existing users already have, don't invent a new one.
Justy: Alright. Download it from conductor dot build if you're on Mac and you're already running agents. Otherwise it's probably not for you yet.
Cody: And thanks for making it to five hundred, Cody.
Justy: That's not how that works.