diff --git a/backend/app/services/llm/enrichment.py b/backend/app/services/llm/enrichment.py
index 39024d5..da4faa8 100644
--- a/backend/app/services/llm/enrichment.py
+++ b/backend/app/services/llm/enrichment.py
@@ -389,8 +389,8 @@ async def enrich_article(article_id: int) -> dict[str, str]:
 
 # === 后台循环 ===
 # 与 translation_loop 一样,常驻从队列里取文章
-ENRICHMENT_INTERVAL_SEC = 5.0  # 没活时等待
-ENRICHMENT_BATCH_SIZE = 3  # 每轮并发拉取候选,然后顺序处理(LLM 客户端本身有节流)
+ENRICHMENT_INTERVAL_SEC = 2.0  # 没活时等待
+ENRICHMENT_BATCH_SIZE = 8  # 每轮并发拉取候选,然后顺序处理(LLM 客户端本身有节流)
 
 
 async def enrichment_loop() -> None:
@@ -438,12 +438,17 @@ async def enrichment_loop() -> None:
                 await asyncio.sleep(ENRICHMENT_INTERVAL_SEC)
                 continue
 
-            for aid in todo_ids:
-                try:
-                    await enrich_article(aid)
-                except Exception as e:
-                    logger.exception("enrich_article %s in loop failed: %s", aid, e)
-                await asyncio.sleep(0.2)  # 文章间轻节流(LLM 内部还有 interval_sec)
+            # 并发 enrich 多篇(LlmClient 内部 interval_sec 已经做了限速,这里只并发不限并发上限)
+            # 但为了不让 Agnes API 同时打太多,加一层并发上限
+            sem = asyncio.Semaphore(3)
+            async def _run_one(aid: int) -> None:
+                async with sem:
+                    try:
+                        await enrich_article(aid)
+                    except Exception as e:
+                        logger.exception("enrich_article %s in loop failed: %s", aid, e)
+
+            await asyncio.gather(*[_run_one(aid) for aid in todo_ids])
         except Exception as e:
             logger.exception("enrichment_loop error: %s", e)
             await asyncio.sleep(ENRICHMENT_INTERVAL_SEC)