What this is
Claude Code this week shipped Custom Agent—essentially letting you write the AI's "persona" into a Markdown file, save it, and later just say "send the reviewer to check this code." We see this as a key step for AI moving from "one-off chat" to "reusable role."
A Juejin tutorial lays out the concepts cleanly: Custom Agent (role template) lives under the .claude/agents/ directory as a Markdown file with YAML frontmatter (configuration metadata); SubAgent is the runtime instance (working employee) spun up from the template each time a task is dispatched; Slash Command (/slash command) is the shortcut executed directly in the main conversation. The three are routinely conflated.
We care more about a design detail that gets glossed over: the tools field—specifying what tools the AI can use. This is not just a capability cap, it's a security mechanism: e.g., a reviewer gets Read (read file) but not Write (edit file), which is far more reliable than writing "don't edit" in the prompt. This detail tells us Anthropic is building toward enterprise deployment, not just catering to geeks.
Industry view
Supporters frame this as a pivotal step for AI productization. The reasoning: turning "roles" into reusable assets mirrors the logic of job descriptions inside enterprises, so IT departments can finally manage AI tasks the way they manage employees; tool permission separation is also foundational infrastructure for enterprise compliance.
But we see at least two risks. First, it still requires users to understand Markdown and basic file structure—the bar hasn't dropped much for non-technical managers. Officially it's "3 minutes to get started," but in practice someone still has to maintain the Agent library long-term. Second, once an Agent library accumulates on one platform, switching costs rise sharply; Cursor, Cline (comparable AI coding tools) already ship similar capabilities, the market landscape is unsettled, and locking in too early means handing over the lifeline.
The deeper problem: most mid-sized domestic enterprises don't even have a standardized process for "AI writing weekly reports," so talking about "Agent library governance" is far-fetched.
Impact on regular people
For mid-sized enterprise IT: you can start breaking down common departmental AI tasks (contract review, report proofreading, code review) into standardized templates invoked uniformly—but you'll need someone to own the configuration metadata and curate the Agent library.
For individual professionals: the productivity gap between people already using Claude Code and those who aren't will keep widening; non-users won't feel it for now, but "knowing how to use AI templates" may be turning into a new baseline skill.
For the consumer market: regular users won't see changes in the short term, but product design thinking is shifting from "chatbot" to "virtual team"—what you buy next may not be an AI assistant, but a dispatchable AI team.