Claude Code Anthropic-style: What changes in practice

On 28 May, XDA published a detailed analysis reproducing step-by-step the Claude Code setup that Anthropic presented at its Code w/ Claude event. This isn't a generic tutorial: the author followed the official architecture precisely and documented what works, what needs adjustment, and where the real bottlenecks are. In that sense, it's more useful than most guides circulating online.

The article's value isn't in discovering anything new, but in validating something many teams suspected: Claude Code's default configuration differs significantly from what Anthropic considers an optimal setup. The gap between "I installed the CLI and it works" and "I have an agent that actually saves time" is bigger than it seems.

What Anthropic's setup implies

Without going into proprietary details, the approach presented at Code w/ Claude pivots around three axes that are already part of Claude Code's public architecture:

• Specialised sub-agents: instead of asking a single agent to do everything, Anthropic's proposal segments tasks—code review, test execution, documentation search—into separate sub-agents that the main agent orchestrates. This reduces error rates in long contexts and makes the workflow more traceable.
• Properly configured MCP servers: the key isn't how many MCP servers are connected, but which ones and with what permissions. XDA's article emphasises that careless configuration of the `claude_desktop_config.json` file can introduce unnecessary calls that slow execution or expose more surface area than needed.
• Hooks at key lifecycle points: hooks like `PreToolUse` and `PostToolUse` allow you to validate or log each tool call before and after it happens. Used wisely, they make Claude Code auditable; ignored, you lose much of the control over what the agent executes locally.

Why it matters and for whom

This configuration is mainly relevant to two profiles. The first: individual developers who already have Claude Code installed but use it as an advanced autocomplete rather than as an autonomous agent. For them, adopting sub-agents and hooks means a paradigm shift with a learning curve, but results in repetitive tasks (test generation, scoped refactoring, automated documentation) are noticeable from the first week.

The second profile is engineering teams evaluating whether Claude Code can integrate into their CI/CD pipeline. Here the sub-agent architecture makes even more sense: each sub-agent can have different permissions and context, making it easier to separate responsibilities without relying on a monolithic prompt that grows out of control.

For occasional users or those who use Claude mainly from the web interface, this configuration offers nothing immediate. The overhead of maintaining MCP servers and hooks isn't worth it unless your work volume is substantial.

What the article doesn't resolve

XDA reproduces the configuration faithfully, but the analysis has a clear limit: it doesn't evaluate the real cost in tokens or the impact on billing when multiple sub-agents are orchestrated in parallel. In projects using Claude Opus 4.7, with its 1M token window, the cost per task can skyrocket if sub-agents carry unnecessary context between calls. This is something any team should measure before scaling this architecture to production.

It also doesn't address credential management when several MCP servers share access to the same resources, a point that in corporate environments often becomes the first real blocker.

---

From ElephantPink, the reading is positive but measured: Anthropic showing reference configurations at public events is useful for normalising best practices, though the gap between a demo and a stable production environment remains each team's responsibility. XDA's article is a good starting point, not a production manual.