DeepSeek's latest open-source code hides a set of numbers: the same conversation, but with a different engineering strategy, sees cache hit rate jump from 6.4% to 59.7%—a 9x gap. This is worth our attention: the biggest cost in deploying AI Agents (AI that can autonomously break down and execute tasks) isn't the model itself, but engineering architecture.
What This Is
This is a DeepSeek-Code source walkthrough on Juejin, covering two core capabilities of Agents: planning mode (think before you act) and sub-agents (dispatch clones to work in parallel).
The article's core insight is called "signal belongs to the model, constraints belong to the harness." The model expresses intent; the system enforces constraints—when entering planning mode, instead of relying on prompts (instructions written for AI) to make the model "read-only," a physical gate is set at the execution layer; the write tool simply won't be callable.
This looks like an engineering detail, yet it delivers a 9x cache hit rate difference. The reason: DeepSeek counts the tool table in the prefix cache (the processed text cached during model invocation—higher hit rate means lower cost). When the tool table changes, the entire conversation history must be re-prefilled (reprocessing all historical text), doubling latency at the most user-sensitive moment.
Industry View
Supporters see this as a signal of Chinese AI companies' engineering maturity—no longer racing on parameters, but starting to race on infrastructure. One practitioner put it bluntly: "Silicon Valley is still comparing benchmark scores; we're already comparing cache hit rates."
But opposing views deserve a hearing. One voice points out that the 9x gap is specific to DeepSeek's implementation (it counts the tool table in the cache hash); OpenAI and Anthropic's protocols may not work the same way—blindly applying this would mislead. A more fundamental critique: point optimization masks the overall complexity of Agents—what really blocks deployment is business process alignment, data compliance, multi-system integration, things these engineering tricks can't solve.
Impact on Regular People
For enterprise IT: when evaluating Agent vendors, model capability is only the entry ticket; engineering architecture (caching strategy, concurrency control, approval circuit breakers) is the long-term cost determinant.
For individual professionals: when running long tasks with AI, "whether it can maintain memory" has a real compute bill behind it—understanding this helps judge which tool is worth paying for.
For consumer markets: AI assistant pricing will continue to diverge—cheapness comes from engineering optimization; goodness isn't just about being smart, but being "smart about saving."