I made a video and posted it with a simple question in the caption: "See how Claude Code uses agents/tasks for parallel tasks (Swarms?)."
The parenthetical question mark was intentional. I genuinely wasn't sure if what I was excited about was a real breakthrough or just more AI doing more things, which might not actually be better.
I've always loved understanding how systems work. Not just using them, but getting into the mechanics of them. So I spent 14 days diving into Claude Code's agent system. Trying to understand what it could actually do and when it mattered.
The demo in the video was deliberately simple and raw. No fancy frameworks. No predefined magic. Just Claude Code in its plain form, and me asking it: "Show me how you handle parallel tasks."
I wanted people to see the mechanism clearly. To understand how the prompting works when you're trying to leverage this capability. Because my hypothesis was that most people wouldn't see it unless it was explained clearly, and that meant showing the actual work, not just the output.
The audio quality wasn't perfect — I'd recorded it in my car on the way to the summer house, which is not an ideal studio setup. But the point came through because the mechanism is genuinely interesting.
Here's what I found: parallel agents are genuinely better for certain kinds of problems. When you have independent tasks, the benefits are real and substantial. Update five API endpoints? Add tests to three modules? Refactor two utility libraries? Parallel agents crush it. You give the orchestrating agent a list of tasks and constraints, and it coordinates multiple worker agents. You get back to what you were doing and come back 30 minutes later to find everything done.
But — and this is important — when tasks are deeply interdependent, parallel agents add overhead. The coordination cost becomes real. You end up waiting for subtask results anyway, and the complexity of managing interdependencies eats the time gain.
So the question in my caption — "(Swarms?)" — became less of a question and more of an observation. Yes, this is swarm-like behavior. But swarms aren't always the right answer. Sometimes serial execution is better. Sometimes you need human judgment in the middle.
The exciting part isn't that Claude Code can do parallel tasks. The exciting part is that it's becoming genuinely difficult to optimize — there are tradeoffs worth thinking about, and you have to make real decisions based on the problem you're solving.
That's a sign of maturity in the tools. When they get good enough that your constraint becomes decision-making rather than capability, you've crossed a threshold.
GitHub Copilot is still playing catch-up here. I posted with hope that they'd get something similar soon. They have the base tools. They have the integration in the IDE. It's not clear why they couldn't offer this kind of capability.
But Anthropic is ahead. And that matters.
The video was a way of saying: if you're using GitHub, watch what Anthropic is doing. Watch how they're thinking about AI that coordinates AI. That's the direction this is going.
🎙️ Hear more about this topic — Parallel agent tasks and swarm orchestration was a key topic on Verbos Podcast #92. → Listen to the episode