diff --git a/README.md b/README.md
index 7fa28ff..b625da9 100644
--- a/README.md
+++ b/README.md
@@ -36,6 +36,13 @@ frontend_port: 4000
 # Max agentic loop iterations (tool call rounds)
 max_iterations: 15
 
+# Sub-agent settings (multi_agent tool)
+sub_agent:
+  max_iterations: 3      # Max tool-call rounds per sub-agent
+  max_tokens: 4096        # Max tokens per LLM call inside a sub-agent
+  max_agents: 5           # Max number of concurrent sub-agents per request
+  max_concurrency: 3      # ThreadPoolExecutor max workers
+
 # Available models
 # Each model must have its own id, name, api_url, api_key
 models:
@@ -117,6 +124,7 @@ backend/
 │       ├── data.py      # 计算器、文本、JSON 处理
 │       ├── weather.py   # 天气查询（模拟）
 │       ├── file_ops.py  # 文件操作（6 个工具，project_id 自动注入）
+│       ├── agent.py     # 多智能体（子 Agent 并发执行，工具权限隔离）
 │       └── code.py      # Python 代码执行（沙箱）
 └── utils/               # 辅助函数
     ├── helpers.py       # 通用函数（ok/err/build_messages 等）
@@ -207,6 +215,7 @@ frontend/
 | **代码执行** | execute_python | 沙箱环境执行 Python |
 | **文件操作** | file_read, file_write, file_delete, file_list, file_exists, file_mkdir | project_id 自动注入 |
 | **天气** | get_weather | 天气查询（模拟） |
+| **智能体** | multi_agent | 派生子 Agent 并发执行（禁止递归，工具权限与主 Agent 一致） |
 
 ## 文档
 
diff --git a/backend/config.py b/backend/config.py
index 232c1ee..bfebf1a 100644
--- a/backend/config.py
+++ b/backend/config.py
@@ -39,3 +39,10 @@ TOOL_MAX_WORKERS = _cfg.get("tool_max_workers", 4)
 
 # Max character length for a single tool result content (truncated if exceeded)
 TOOL_RESULT_MAX_LENGTH = _cfg.get("tool_result_max_length", 4096)
+
+# Sub-agent settings (multi_agent tool)
+_sa = _cfg.get("sub_agent", {})
+SUB_AGENT_MAX_ITERATIONS = _sa.get("max_iterations", 3)
+SUB_AGENT_MAX_TOKENS = _sa.get("max_tokens", 4096)
+SUB_AGENT_MAX_AGENTS = _sa.get("max_agents", 5)
+SUB_AGENT_MAX_CONCURRENCY = _sa.get("max_concurrency", 3)
diff --git a/backend/tools/builtin/agent.py b/backend/tools/builtin/agent.py
index 0b3f7a1..503a8ef 100644
--- a/backend/tools/builtin/agent.py
+++ b/backend/tools/builtin/agent.py
@@ -4,12 +4,25 @@ Provides:
 - multi_agent: Spawn sub-agents with independent LLM conversation loops
 """
 import json
+import logging
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from typing import List, Dict, Any, Optional
 
 from backend.tools.factory import tool
 from backend.tools.core import registry
 from backend.tools.executor import ToolExecutor
+from backend.config import (
+    DEFAULT_MODEL,
+    SUB_AGENT_MAX_ITERATIONS,
+    SUB_AGENT_MAX_TOKENS,
+    SUB_AGENT_MAX_AGENTS,
+    SUB_AGENT_MAX_CONCURRENCY,
+)
+
+logger = logging.getLogger(__name__)
+
+# Sub-agents are forbidden from using multi_agent to prevent infinite recursion
+BLOCKED_TOOLS = {"multi_agent"}
 
 
 def _to_executor_calls(tool_calls: list, id_prefix: str = "tc") -> list:
@@ -68,13 +81,16 @@ def _run_sub_agent(
             "error": "LLM client not available",
         }
 
-    # Build tool list – filter to requested tools or use all
+    # Build tool list – filter to requested tools, then remove blocked
     all_tools = registry.list_all()
     if tool_names:
         allowed = set(tool_names)
         tools = [t for t in all_tools if t["function"]["name"] in allowed]
     else:
-        tools = all_tools
+        tools = list(all_tools)
+
+    # Remove blocked tools to prevent recursion
+    tools = [t for t in tools if t["function"]["name"] not in BLOCKED_TOOLS]
 
     executor = ToolExecutor(registry=registry)
     context = {"model": model}
@@ -128,7 +144,17 @@ def _run_sub_agent(
                 })
                 tc_list = message["tool_calls"]
                 executor_calls = _to_executor_calls(tc_list)
-                tool_results = executor.process_tool_calls(executor_calls, context)
+                # Execute tools inside app_context – file ops and other DB-
+                # dependent tools require an active Flask context and session.
+                with app.app_context():
+                    if len(executor_calls) > 1:
+                        tool_results = executor.process_tool_calls_parallel(
+                            executor_calls, context
+                        )
+                    else:
+                        tool_results = executor.process_tool_calls(
+                            executor_calls, context
+                        )
                 messages.extend(tool_results)
             else:
                 # Final text response
@@ -159,7 +185,7 @@ def _run_sub_agent(
         "Spawn multiple sub-agents to work on tasks concurrently. "
         "Each agent runs its own independent conversation with the LLM and can use tools. "
         "Useful for parallel research, multi-file analysis, or dividing complex tasks into sub-tasks. "
-        "Each agent is limited to 3 iterations and 4096 tokens to control cost."
+        "Resource limits (iterations, tokens, concurrency) are configured in config.yml -> sub_agent."
     ),
     parameters={
         "type": "object",
@@ -221,19 +247,18 @@ def multi_agent(arguments: dict) -> dict:
 
     tasks = arguments["tasks"]
 
-    if len(tasks) > 5:
-        return {"success": False, "error": "Maximum 5 concurrent agents allowed"}
+    if len(tasks) > SUB_AGENT_MAX_AGENTS:
+        return {"success": False, "error": f"Maximum {SUB_AGENT_MAX_AGENTS} concurrent agents allowed"}
 
     # Get current conversation context for model/project info
     app = current_app._get_current_object()
 
     # Use injected model/project_id from executor context, fall back to defaults
-    from backend.config import DEFAULT_MODEL
     model = arguments.get("_model") or DEFAULT_MODEL
     project_id = arguments.get("_project_id")
 
-    # Execute agents concurrently (max 3 at a time)
-    concurrency = min(len(tasks), 3)
+    # Execute agents concurrently
+    concurrency = min(len(tasks), SUB_AGENT_MAX_CONCURRENCY)
     results = [None] * len(tasks)
 
     with ThreadPoolExecutor(max_workers=concurrency) as pool:
@@ -244,9 +269,10 @@ def multi_agent(arguments: dict) -> dict:
                 task["instruction"],
                 task.get("tools"),
                 model,
-                4096,
+                SUB_AGENT_MAX_TOKENS,
                 project_id,
                 app,
+                SUB_AGENT_MAX_ITERATIONS,
             ): i
             for i, task in enumerate(tasks)
         }
diff --git a/backend/utils/helpers.py b/backend/utils/helpers.py
index 6691002..2acb3c2 100644
--- a/backend/utils/helpers.py
+++ b/backend/utils/helpers.py
@@ -1,6 +1,6 @@
 """Common helper functions"""
 import json
-from datetime import date, datetime
+from datetime import date, datetime, timezone
 from typing import Any
 from flask import jsonify
 from backend import db
@@ -97,7 +97,7 @@ def message_to_dict(msg: Message) -> dict:
 
 def record_token_usage(user_id, model, prompt_tokens, completion_tokens):
     """Record token usage"""
-    today = date.today()
+    today = datetime.now(timezone.utc).date()
     usage = TokenUsage.query.filter_by(
         user_id=user_id, date=today, model=model
     ).first()
diff --git a/docs/Design.md b/docs/Design.md
index 0647108..fe16f1a 100644
--- a/docs/Design.md
+++ b/docs/Design.md
@@ -66,6 +66,7 @@ backend/
 │       ├── data.py      # 计算器、文本、JSON
 │       ├── weather.py   # 天气查询
 │       ├── file_ops.py  # 文件操作（project_id 自动注入）
+│       ├── agent.py     # 多智能体（子 Agent 并发执行，工具权限隔离）
 │       └── code.py      # 代码执行
 │
 ├── utils/               # 辅助函数
@@ -1020,6 +1021,13 @@ frontend_port: 4000
 # 智能体循环最大迭代次数（工具调用轮次上限，默认 5）
 max_iterations: 15
 
+# 子代理资源配置（multi_agent 工具）
+sub_agent:
+  max_iterations: 3      # 每个子代理的最大工具调用轮数
+  max_tokens: 4096        # 每次调用的最大 token 数
+  max_agents: 5           # 每次请求最多派生的子代理数
+  max_concurrency: 3      # 并发线程数
+
 # 可用模型列表（每个模型必须指定 api_url 和 api_key）
 # 支持任何 OpenAI 兼容 API（DeepSeek、GLM、OpenAI、Moonshot、Qwen 等）
 models:
diff --git a/docs/ToolSystemDesign.md b/docs/ToolSystemDesign.md
index 1c76288..f56df60 100644
--- a/docs/ToolSystemDesign.md
+++ b/docs/ToolSystemDesign.md
@@ -249,14 +249,25 @@ file_read({"path": "src/main.py", "project_id": "xxx"})
 
 | 工具名称 | 描述 | 参数 |
 |---------|------|------|
-| `multi_agent` | 派生子 Agent 并发执行任务（最多 5 个） | `tasks`: 任务数组（name, instruction, tools）<br>`_model`: 模型名称（自动注入）<br>`_project_id`: 项目 ID（自动注入） |
+| `multi_agent` | 派生子 Agent 并发执行任务 | `tasks`: 任务数组（name, instruction, tools）<br>`_model`: 模型名称（自动注入）<br>`_project_id`: 项目 ID（自动注入） |
 
 **`multi_agent` 工作原理：**
 1. 接收任务数组，每个任务指定 name、instruction 和可选的 tools 列表
-2. 为每个子 Agent 创建独立线程，各自拥有 LLM 对话循环（最多 3 轮迭代，4096 tokens）
-3. 通过 Service Locator 获取 `llm_client` 实例
-4. 子 Agent 在 `app.app_context()` 中运行，可独立调用所有注册工具
-5. 返回 `{success, results: [{task_name, success, response/error}], total}`
+2. 子 Agent **禁止使用 `multi_agent` 工具**（`BLOCKED_TOOLS`），防止无限递归
+3. 子 Agent 工具权限与主 Agent 一致（除 multi_agent 外的所有已注册工具），支持并行工具执行
+4. 为每个子 Agent 创建独立线程，各自拥有 LLM 对话循环
+5. 子 Agent 在 `app.app_context()` 中运行 LLM 调用和工具执行，确保数据库等依赖正常工作
+6. 通过 Service Locator 获取 `llm_client` 实例
+7. 返回 `{success, results: [{task_name, success, response/error}], total}`
+
+**资源配置**（`config.yml` → `sub_agent`）：
+
+| 配置项 | 默认值 | 说明 |
+|--------|--------|------|
+| `max_iterations` | 3 | 每个子代理的最大工具调用轮数 |
+| `max_tokens` | 4096 | 每次调用的最大 token 数 |
+| `max_agents` | 5 | 每次请求最多派生的子代理数 |
+| `max_concurrency` | 3 | ThreadPoolExecutor 并发线程数 |
 
 ---
 
diff --git a/frontend/src/components/StatsPanel.vue b/frontend/src/components/StatsPanel.vue
index 66ea9f4..7d640d1 100644
--- a/frontend/src/components/StatsPanel.vue
+++ b/frontend/src/components/StatsPanel.vue
@@ -132,8 +132,15 @@ const sortedDaily = computed(() => {
 const chartData = computed(() => {
   if (period.value === 'daily' && stats.value?.hourly) {
     const hourly = stats.value.hourly
+    // Backend returns UTC hours — convert to local timezone for display.
+    const offset = -new Date().getTimezoneOffset() / 60  // e.g. +8 for UTC+8
+    const localHourly = {}
+    for (const [utcH, val] of Object.entries(hourly)) {
+      const localH = ((parseInt(utcH) + offset) % 24 + 24) % 24
+      localHourly[localH] = val
+    }
     let minH = 24, maxH = -1
-    for (const h of Object.keys(hourly)) {
+    for (const h of Object.keys(localHourly)) {
       const hour = parseInt(h)
       if (hour < minH) minH = hour
       if (hour > maxH) maxH = hour
@@ -145,16 +152,19 @@ const chartData = computed(() => {
       const h = start + i
       return {
         label: `${h}:00`,
-        value: hourly[String(h)]?.total || 0,
+        value: localHourly[String(h)]?.total || 0,
       }
     })
   }
 
   const data = sortedDaily.value
   return Object.entries(data).map(([date, val]) => {
-    const d = new Date(date)
+    // date is "YYYY-MM-DD" from backend — parse directly to avoid
+    // new Date() timezone shift (parsed as UTC midnight then
+    // getMonth/getDate applies local offset, potentially off by one day).
+    const [year, month, day] = date.split('-')
     return {
-      label: `${d.getMonth() + 1}/${d.getDate()}`,
+      label: `${parseInt(month)}/${parseInt(day)}`,
       value: val.total,
       prompt: val.prompt || 0,
       completion: val.completion || 0,