If you’ve been using AI coding tools for the past year, you’ve probably noticed a recurring bottleneck: the waiting game. You ask your agent to investigate a deeply buried bug or scaffold a new microservice, and then you’re stuck staring at your terminal. You can’t really move forward until the AI finishes mapping the codebase and churning through files. Anthropic is about to change that dynamic entirely. The upcoming version of Claude Code is flipping the script by pushing its subagent workflows into the background by default. Instead of locking up your main session, the system will now let you keep talking, coding, and planning while your delegated AI tasks run quietly in parallel.

The feature has been quietly rolling out to a small subset of users over the past week, but the official word came directly from Anthropic’s Boris Cherny. In a recent post on X, Cherny confirmed the shift, noting that if developers actually want an agent to block the main thread and run in the foreground, they’ll now have to explicitly tell Claude to do so. It’s a massive quality-of-life upgrade for engineers who treat their terminal as a command center rather than a simple chat interface. By shifting to a background-first model, Claude Code is acting less like a junior developer you have to handhold and more like a capable async team.

Of course, letting AI agents run unseen in the background is a concept that rightfully makes some developers nervous. The idea of an AI wandering off into your repository while you aren’t looking sounds like a recipe for a bad git commit. But Anthropic seems to have anticipated the trust issues. Cherny clarified that any time a background subagent needs to perform an action requiring approval, it forwards the permission request directly to your main agent session. You’re never entirely out of the loop. If you want to check in on exactly what a subagent is doing, you can simply press the down arrow and hit enter to “zoom in” on that specific instance, message it directly, and then step back out. A dedicated, separate tab interface for managing these parallel workers is also on the way, though it’s already accessible for users who have opted into the agent teams feature.

To understand why this is such a big deal, you have to look at how Claude Code fundamentally handles complex architecture. When you ask the main agent to do something massive—like auditing security across an entire monorepo—loading all that data into a single context window is a disaster. It burns through your token budget and clutters the conversation with noise. Subagents were introduced to solve this. They are isolated, single-purpose AI instances with their own prompt and tool access. You tell the main agent to figure something out, and it spins up a subagent to go read fifty files, summarize the findings, and bring back only the final answer. It’s an incredibly elegant architecture, but until now, it still operated sequentially from the user’s perspective.

With background execution, the ceiling for what you can accomplish in a single coding sprint gets significantly higher. Developers can now fire off multiple parallel investigations—say, having one subagent map out data dependencies while another writes test coverage for a new component—and continue refining the core logic on the main thread. Interestingly, Cherny also confirmed that this delegation can get remarkably deep. These background subagents are already capable of spawning their own subagents up to five layers deep. It’s a fractal approach to problem-solving, where an investigation task can fan out across your codebase, delegating smaller checks to specialized workers, all while you focus on the big picture.

It’s not entirely without friction, though. The developer community is already pointing out some edge cases that come with this new async reality. When tasks fan out and run in the background, cache invalidation becomes a real headache. A standard five-minute cache might expire while you’re waiting for a complex, multi-layered subagent workflow to report back, leading to inefficiencies and higher token costs. There are also growing demands for more customized control, with engineers asking for the ability to save, modify, and load their own JSON-based workflows with specific fan-out loops, as well as better ways to review the resulting code directly inside the IDE. Anthropic is apparently cooking up solutions for the IDE review process, but managing state and context across long-running background tasks will likely be an ongoing challenge as these tools scale.

Ultimately, this update represents a crucial shift in how we interact with AI development tools. We are moving away from the paradigm of the AI as a hyper-fast typist and moving toward the AI as a project manager coordinating its own tiny workforce. By pushing subagents into the background, Claude Code is giving engineers the space to do what they actually do best: orchestrate, design, and make decisions, rather than just babysitting an algorithmic assistant. It’s a glimpse into the near future of software development, where your terminal is always humming with parallel activity, and your only real bottleneck is how fast you can think up the next task.