[Architecture] The Subprocess-based Agent Loop causes poor performance and memory bottlenecks in server-side/high-concurrency scenarios

**1. Spawning a new subprocess for every new session**

The current architectural design of claude-agent-sdk-typescript suffers from severe concurrency and performance bottlenecks. Specifically, the SDK spawns a new isolated subprocess for every single new session. While this tightly coupled Agent Loop architecture might be reasonable for standalone CLI tools, it is disastrous for server-side applications.

**2. Claude Agent SDK vs LangChain DeepAgent**

**Runaway Resource Exhaustion:**
As concurrency scales from 10 to 40, the Claude SDK's CPU usage skyrockets from 27% to 99%, with memory spiking directly to 86%. In stark contrast, the DeepAgent architecture (which completely avoids the subprocess overhead) maintains a mere 15% CPU usage and a rock-solid 8% memory footprint. This proves that the Claude SDK's compute and memory resources are entirely cannibalized by meaningless process management.

**Deadlocked Throughput Bottleneck:**
Upon reaching a concurrency of 40, the Claude SDK's extreme system load causes its QPM (Queries Per Minute) to completely stall at a mere 28.8, while the P95 latency degrades to an unacceptable 103 seconds.

**Extremely Low Concurrency Ceiling:**
The data clearly demonstrates that at a mere 40 concurrent requests, the Claude SDK entirely drains all resources of a 4C8G server, rendering it completely unresponsive and unable to provide service beyond that point. Meanwhile, a well-designed architecture effortlessly handles up to 100 concurrent requests with plenty of headroom (utilizing only 37% CPU).

**Hoping for an official resolution, as this currently makes the SDK unusable in production server environments.**