AI 智能體有 93% token 預算被浪費：原因何在
93% of your AI agent's token budget is going straight to waste. Here is why.
Miracuves 指出，AI 智能體上下文窗口中超過 90% 的 token 是模型已處理過的系統提示詞、歷史狀態及工具輸出。Miracuves 以語義狀態分層和增量上下文修剪，減少重複內容，只傳遞狀態變化及工作記憶，並在工作流程中清除過時的工具輸出。該方案帶來低於一秒的執行延遲、減少 API 開支，並稱長流程智能體循環中上下文沒有退化。
Your AI agents are likely choking on their own memory.
When benchmarking a complex, multi-step agent workflow recently, the telemetry revealed an uncomfortable truth: over 90% of the tokens being pumped into the context window were redundant payloads—system prompts, historical states, and intermediate tool outputs the model had already processed multiple times.
If you are building production-grade agents, throwing larger context windows at this problem is just burning cash. The real bottleneck is architectural: most agent frameworks treat state as a monolithic, brute-force append-only log.
Here is how we tackled this exact problem at Miracuves by engineering a Semantic State Layering & Incremental Context Pruning mechanism:
State Decoupling: Instead of re-sending the entire conversation history on every execution step, Miracuves isolates static system instructions and historical milestones into an optimized reference layer.
Dynamic Delta Caching: Only state changes and active working memory are passed downstream to the LLM, cutting redundant payload size drastically.
Smart Context Pruning: Automatically garbage-collects obsolete intermediate tool outputs mid-workflow before they pollute the reasoning context.
The result? Sub-second execution latencies, slashed API bills, and zero context degradation during long-horizon agent loops.
How are you handling state management and context bloat in your agent pipelines right now? Are you rolling your own caching layers or relying purely on provider-level prompt caching? Let's debate in the comments.
來源：DEV Community · dev.to