diff --git a/src/agentsight/dashboard/src/i18n.tsx b/src/agentsight/dashboard/src/i18n.tsx
index 40c497638..139d4445a 100644
--- a/src/agentsight/dashboard/src/i18n.tsx
+++ b/src/agentsight/dashboard/src/i18n.tsx
@@ -2,8 +2,10 @@ import React, {
   createContext,
   useCallback,
   useContext,
+  useEffect,
   useLayoutEffect,
   useMemo,
+  useRef,
   useState,
 } from 'react';
 
@@ -35,6 +37,19 @@ const enUSMessages = {
   'nav.riskEnforcement': 'Risk Enforcement',
   'nav.trajectoryViewer': 'Trajectory Viewer',
   'nav.settings': 'Settings',
+  'latency.title': 'Latency Metrics',
+  'latency.ttft': 'TTFT',
+  'latency.tps': 'TPS',
+  'latency.tpot': 'TPOT',
+  'latency.e2e': 'E2E',
+  'latency.p50': 'P50',
+  'latency.p95': 'P95',
+  'latency.p99': 'P99',
+  'latency.loading': 'Loading latency metrics...',
+  'latency.error': 'Failed to load latency metrics',
+  'latency.range24h': 'Last 24h',
+  'latency.range7d': 'Last 7d',
+  'latency.range30d': 'Last 30d',
   'login.subtitle': 'Enter your dashboard token to continue',
   'login.tokenLabel': 'Dashboard Token',
   'login.tokenPlaceholder': 'Paste your token here',
@@ -1016,6 +1031,19 @@ export const messages: Record<Locale, Record<MessageKey, string>> = {
     'nav.riskEnforcement': '风险拦截',
     'nav.trajectoryViewer': '轨迹查看',
     'nav.settings': '设置',
+    'latency.title': '延迟指标',
+    'latency.ttft': 'TTFT',
+    'latency.tps': 'TPS',
+    'latency.tpot': 'TPOT',
+    'latency.e2e': 'E2E',
+    'latency.p50': 'P50',
+    'latency.p95': 'P95',
+    'latency.p99': 'P99',
+    'latency.loading': '正在加载延迟指标...',
+    'latency.error': '延迟指标加载失败',
+    'latency.range24h': '最近 24h',
+    'latency.range7d': '最近 7d',
+    'latency.range30d': '最近 30d',
     'login.subtitle': '请输入 Dashboard 令牌以继续',
     'login.tokenLabel': 'Dashboard 令牌',
     'login.tokenPlaceholder': '在此粘贴令牌',
@@ -2135,33 +2163,87 @@ interface LanguageSwitcherProps {
   className?: string;
 }
 
+const LOCALE_OPTIONS: Array<{ value: Locale; label: string }> = [
+  { value: 'en-US', label: 'English' },
+  { value: 'zh-CN', label: '简体中文' },
+];
+
 export const LanguageSwitcher: React.FC<LanguageSwitcherProps> = ({
   id,
   className = '',
 }) => {
   const { locale, setLocale, t } = useI18n();
+  const [open, setOpen] = useState(false);
+  const containerRef = useRef<HTMLDivElement>(null);
 
-  const handleChange = (event: React.ChangeEvent<HTMLSelectElement>) => {
-    const nextLocale = event.target.value;
-    if (isSupportedLocale(nextLocale)) {
-      setLocale(nextLocale);
-    }
-  };
+  useEffect(() => {
+    if (!open) return;
+
+    const closeOnPointerDown = (event: PointerEvent) => {
+      if (!containerRef.current?.contains(event.target as Node)) {
+        setOpen(false);
+      }
+    };
+    const closeOnEscape = (event: KeyboardEvent) => {
+      if (event.key === 'Escape') {
+        setOpen(false);
+      }
+    };
+
+    document.addEventListener('pointerdown', closeOnPointerDown);
+    document.addEventListener('keydown', closeOnEscape);
+    return () => {
+      document.removeEventListener('pointerdown', closeOnPointerDown);
+      document.removeEventListener('keydown', closeOnEscape);
+    };
+  }, [open]);
+
+  const selectedLabel = LOCALE_OPTIONS.find((option) => option.value === locale)?.label ?? 'English';
 
   return (
-    <label className={`inline-flex items-center gap-1.5 ${className}`} htmlFor={id}>
+    <div ref={containerRef} className={`relative inline-flex items-center gap-1.5 ${className}`}>
       <span aria-hidden="true">🌐</span>
-      <span className="sr-only">{t('language.label')}</span>
-      <select
+      <button
+        type="button"
         id={id}
-        aria-label={t('language.label')}
-        value={locale}
-        onChange={handleChange}
-        className="rounded-md border border-gray-300 bg-white px-2 py-1.5 text-sm text-gray-700 focus:outline-none focus:ring-2 focus:ring-blue-500"
+        aria-label={`${t('language.label')}: ${selectedLabel}`}
+        aria-haspopup="menu"
+        aria-expanded={open}
+        aria-controls={`${id}-menu`}
+        onClick={() => setOpen((current) => !current)}
+        className="inline-flex items-center gap-1 rounded-md border border-gray-300 bg-white px-2 py-1.5 text-sm text-gray-700 focus:outline-none focus:ring-2 focus:ring-blue-500"
       >
-        <option value="en-US">English</option>
-        <option value="zh-CN">简体中文</option>
-      </select>
-    </label>
+        {selectedLabel}
+        <span aria-hidden="true" className="text-xs text-gray-400">▾</span>
+      </button>
+      {open && (
+        <div
+          id={`${id}-menu`}
+          role="menu"
+          aria-label={t('language.label')}
+          className="absolute right-0 top-full z-50 mt-1 min-w-[120px] rounded-md border border-gray-200 bg-white p-1 shadow-lg"
+        >
+          {LOCALE_OPTIONS.map((option) => (
+            <button
+              key={option.value}
+              type="button"
+              role="menuitemradio"
+              aria-checked={locale === option.value}
+              onClick={() => {
+                setLocale(option.value);
+                setOpen(false);
+              }}
+              className={`block w-full rounded px-3 py-2 text-left text-sm ${
+                locale === option.value
+                  ? 'bg-blue-50 text-blue-700'
+                  : 'text-gray-700 hover:bg-gray-100'
+              }`}
+            >
+              {option.label}
+            </button>
+          ))}
+        </div>
+      )}
+    </div>
   );
 };
diff --git a/src/agentsight/dashboard/src/pages/AgentHealthPage.tsx b/src/agentsight/dashboard/src/pages/AgentHealthPage.tsx
index f5b724942..cdea3f15c 100644
--- a/src/agentsight/dashboard/src/pages/AgentHealthPage.tsx
+++ b/src/agentsight/dashboard/src/pages/AgentHealthPage.tsx
@@ -5,8 +5,9 @@ import {
   restartAgentHealth,
   fetchInterruptions,
   resolveInterruption,
+  fetchLatencyMetrics,
 } from '../utils/apiClient';
-import type { InterruptionRecord, InterruptionSeverity } from '../utils/apiClient';
+import type { InterruptionRecord, InterruptionSeverity, LatencyMetricsSummary, MetricPercentiles } from '../utils/apiClient';
 import type { AgentHealthStatus } from '../types';
 import { useI18n, useLocaleTag, INTERRUPTION_TYPES, interruptionTypeKey } from '../i18n';
 import type { MessageKey } from '../i18n';
@@ -59,13 +60,85 @@ interface Toast {
   message: string;
 }
 
+function formatMetricValue(value: number): string {
+  return value.toLocaleString(undefined, { maximumFractionDigits: 2 });
+}
+
+function formatMetricP50(metric: MetricPercentiles | null, unit: string): string {
+  return metric ? formatMetricValue(metric.p50) + ' ' + unit : '—';
+}
+
+function canonicalAgentKey(agentName: string): string {
+  return agentName.toLowerCase();
+}
+
+const LatencyMetricsRow: React.FC<{ metrics: LatencyMetricsSummary }> = ({ metrics }) => {
+  const { t } = useI18n();
+  const items = [
+    { label: t('latency.ttft'), metric: metrics.ttft_ms, unit: 'ms' },
+    { label: t('latency.tps'), metric: metrics.tps_tokens_per_second, unit: 'tokens/s' },
+    { label: t('latency.tpot'), metric: metrics.tpot_ms_per_token, unit: 'ms/token' },
+    { label: t('latency.e2e'), metric: metrics.e2e_latency_ms, unit: 'ms' },
+  ];
+
+  if (!items.some(item => item.metric !== null)) return null;
+
+  const tooltip = items
+    .map(({ label, metric, unit }) => {
+      if (!metric) return label + ' —';
+      return (
+        label +
+        ' ' +
+        t('latency.p50') +
+        ' ' +
+        formatMetricValue(metric.p50) +
+        ' ' +
+        unit +
+        ' · ' +
+        t('latency.p95') +
+        ' ' +
+        formatMetricValue(metric.p95) +
+        ' ' +
+        unit +
+        ' · ' +
+        t('latency.p99') +
+        ' ' +
+        formatMetricValue(metric.p99) +
+        ' ' +
+        unit
+      );
+    })
+    .join(' · ');
+
+  return (
+    <div
+      className="mt-2 pt-2 border-t border-gray-100 flex flex-wrap gap-x-3 gap-y-1 text-[11px] text-gray-700"
+      title={tooltip}
+      aria-label={tooltip}
+    >
+      {items.map(({ label, metric, unit }) => (
+        <span key={label}>
+          <span className="text-gray-400">{label}</span> {formatMetricP50(metric, unit)}
+        </span>
+      ))}
+    </div>
+  );
+};
+
+const LATENCY_TIME_PRESETS = [
+  { key: 'latency.range24h', ms: 24 * 3600 * 1000 },
+  { key: 'latency.range7d', ms: 7 * 24 * 3600 * 1000 },
+  { key: 'latency.range30d', ms: 30 * 24 * 3600 * 1000 },
+] as const;
+
 const AgentCard: React.FC<{
   agent: AgentHealthStatus;
   related: AgentHealthStatus[];
   onDelete: (pid: number) => void;
   onRestart: (pid: number) => void;
   restarting: boolean;
-}> = ({ agent, related, onDelete, onRestart, restarting }) => {
+  latency?: LatencyMetricsSummary;
+}> = ({ agent, related, onDelete, onRestart, restarting, latency }) => {
   const { t } = useI18n();
   const [showRelated, setShowRelated] = useState(false);
 
@@ -177,6 +250,7 @@ const AgentCard: React.FC<{
           </div>
         )}
       </div>
+      {latency && <LatencyMetricsRow metrics={latency} />}
       {(isOffline || canRestart) && (
         <div className="mt-2 flex items-center gap-3">
           {isOffline && (
@@ -236,6 +310,34 @@ const AgentStatusSection: React.FC<{ addToast: (msg: string) => void }> = ({ add
   const [error, setError] = useState<string | null>(null);
   const [restartingPids, setRestartingPids] = useState<Set<number>>(new Set());
   const hasDataRef = useRef(false);
+  const [rangeMs, setRangeMs] = useState(7 * 24 * 3600 * 1000);
+  const [latencyMetrics, setLatencyMetrics] = useState<LatencyMetricsSummary[]>([]);
+  const [latencyLoading, setLatencyLoading] = useState(true);
+  const [latencyError, setLatencyError] = useState<string | null>(null);
+  const latencyRequestIdRef = useRef(0);
+
+  const loadLatency = useCallback(async () => {
+    const requestId = ++latencyRequestIdRef.current;
+    setLatencyLoading(true);
+    setLatencyError(null);
+    try {
+      const endNs = Date.now() * 1_000_000;
+      const startNs = endNs - rangeMs * 1_000_000;
+      const data = await fetchLatencyMetrics(startNs, endNs);
+      if (requestId === latencyRequestIdRef.current) {
+        setLatencyMetrics(data);
+        setLatencyError(null);
+      }
+    } catch (e: any) {
+      if (requestId === latencyRequestIdRef.current) {
+        setLatencyError(e.message || '');
+      }
+    } finally {
+      if (requestId === latencyRequestIdRef.current) {
+        setLatencyLoading(false);
+      }
+    }
+  }, [rangeMs]);
 
   const refresh = useCallback(async () => {
     try {
@@ -289,6 +391,10 @@ const AgentStatusSection: React.FC<{ addToast: (msg: string) => void }> = ({ add
     return () => clearInterval(timer);
   }, [refresh]);
 
+  useEffect(() => {
+    void loadLatency();
+  }, [loadLatency]);
+
   // Sort: hung/unhealthy first (real problems), healthy in the middle, offline last (less prominent)
   const sorted = [...agents].sort((a, b) => {
     const order: Record<string, number> = {
@@ -307,6 +413,24 @@ const AgentStatusSection: React.FC<{ addToast: (msg: string) => void }> = ({ add
   const hungCount = agents.filter(a => a.status === 'hung').length;
   const totalCount = agents.length;
 
+  const latencyByAgent = new Map<string, LatencyMetricsSummary[]>();
+  for (const metric of latencyMetrics) {
+    if (metric.agent_name !== null) {
+      const key = canonicalAgentKey(metric.agent_name);
+      const summaries = latencyByAgent.get(key);
+      if (summaries) {
+        summaries.push(metric);
+      } else {
+        latencyByAgent.set(key, [metric]);
+      }
+    }
+  }
+  const latencyForAgent = (agentName: string): LatencyMetricsSummary | undefined => {
+    const summaries = latencyByAgent.get(canonicalAgentKey(agentName));
+    // Do not silently choose one when casing variants produce separate summaries.
+    return summaries?.length === 1 ? summaries[0] : undefined;
+  };
+
   const gatewayPids = new Set(sorted.map(a => a.pid));
   // Orphaned related processes: workers whose parent is not any main card (should not happen, fallback).
   // Filter out status=offline processes — they are TTL-cleaned after 5 minutes.
@@ -342,9 +466,35 @@ const AgentStatusSection: React.FC<{ addToast: (msg: string) => void }> = ({ add
             </span>
           )}
         </div>
-        {lastScan > 0 && (
-          <span className="text-xs text-gray-400">{t('ah.lastScan', { time: relativeTime(lastScan, t) })}</span>
-        )}
+        <div className="flex items-center gap-2 flex-wrap justify-end">
+          <div className="flex items-center gap-1.5">
+            <span className="text-[11px] text-gray-400">{t('latency.title')}</span>
+            {LATENCY_TIME_PRESETS.map(({ key, ms }) => (
+              <button
+                key={key}
+                type="button"
+                onClick={() => setRangeMs(ms)}
+                aria-pressed={rangeMs === ms}
+                className={rangeMs === ms
+                  ? 'px-2 py-1 text-[11px] rounded bg-blue-100 text-blue-700 font-medium'
+                  : 'px-2 py-1 text-[11px] rounded bg-gray-100 hover:bg-gray-200 text-gray-600'}
+              >
+                {t(key)}
+              </button>
+            ))}
+          </div>
+          {latencyLoading && (
+            <span className="text-[11px] text-gray-400">{t('latency.loading')}</span>
+          )}
+          {latencyError !== null && (
+            <span className="text-[11px] text-red-400" title={latencyError || t('latency.error')}>
+              {t('latency.error')}
+            </span>
+          )}
+          {lastScan > 0 && (
+            <span className="text-xs text-gray-400">{t('ah.lastScan', { time: relativeTime(lastScan, t) })}</span>
+          )}
+        </div>
       </div>
 
       {loading ? (
@@ -367,6 +517,7 @@ const AgentStatusSection: React.FC<{ addToast: (msg: string) => void }> = ({ add
               onDelete={handleDelete}
               onRestart={handleRestart}
               restarting={restartingPids.has(agent.pid)}
+              latency={latencyForAgent(agent.agent_name)}
             />
           ))}
         </div>
diff --git a/src/agentsight/dashboard/src/utils/apiClient.ts b/src/agentsight/dashboard/src/utils/apiClient.ts
index ee67d46e8..4cfb63558 100644
--- a/src/agentsight/dashboard/src/utils/apiClient.ts
+++ b/src/agentsight/dashboard/src/utils/apiClient.ts
@@ -557,6 +557,38 @@ export interface TimeseriesResponse {
   model_series: ModelTimeseriesBucket[];
 }
 
+export interface MetricPercentiles {
+  p50: number;
+  p95: number;
+  p99: number;
+}
+
+export interface LatencyMetricsSummary {
+  agent_name: string | null;
+  call_count: number;
+  streaming_call_count: number;
+  ttft_ms: MetricPercentiles | null;
+  tps_tokens_per_second: MetricPercentiles | null;
+  tpot_ms_per_token: MetricPercentiles | null;
+  e2e_latency_ms: MetricPercentiles | null;
+}
+
+/**
+ * Fetch percentile latency metrics grouped by Agent.
+ */
+export async function fetchLatencyMetrics(
+  startNs: number,
+  endNs: number,
+  agentName?: string
+): Promise<LatencyMetricsSummary[]> {
+  const params = new URLSearchParams({
+    start_ns: String(startNs),
+    end_ns: String(endNs),
+  });
+  if (agentName) params.set('agent_name', agentName);
+  return apiFetch<LatencyMetricsSummary[]>(`${API_BASE}/api/metrics/latency?${params.toString()}`);
+}
+
 /**
  * Fetch time-bucketed token stats and per-model breakdowns.
  */
diff --git a/src/agentsight/dashboard/tests/apiClient-regression.test.cjs b/src/agentsight/dashboard/tests/apiClient-regression.test.cjs
index ebac6d11e..7a7d5bfb0 100644
--- a/src/agentsight/dashboard/tests/apiClient-regression.test.cjs
+++ b/src/agentsight/dashboard/tests/apiClient-regression.test.cjs
@@ -8,6 +8,7 @@ const {
   enforcementSupportsMode,
   enforcementViolationTotal,
   fetchContainmentPlan,
+  fetchLatencyMetrics,
   fetchSecurityCase,
   fetchSecurityStatus,
   reviewSecurityCase,
@@ -114,6 +115,49 @@ test('fetchSecurityStatus preserves a non-2xx availability state envelope', asyn
   assert.deepEqual(response.data, { error: 'socket unavailable' });
 });
 
+test('fetchLatencyMetrics forwards ranges and preserves nullable percentile data', async () => {
+  let requestedUrl = null;
+  global.fetch = async (url) => {
+    requestedUrl = String(url);
+    return new Response(JSON.stringify([{
+      agent_name: 'claude',
+      call_count: 3,
+      streaming_call_count: 2,
+      ttft_ms: { p50: 10, p95: 20, p99: 30 },
+      tps_tokens_per_second: { p50: 40, p95: 50, p99: 60 },
+      tpot_ms_per_token: null,
+      e2e_latency_ms: { p50: 100, p95: 200, p99: 300 },
+    }]), { status: 200 });
+  };
+
+  const response = await fetchLatencyMetrics(1_000_000_000, 2_000_000_000, 'claude');
+  const url = new URL(requestedUrl);
+
+  assert.equal(url.pathname, '/api/metrics/latency');
+  assert.equal(url.searchParams.get('start_ns'), '1000000000');
+  assert.equal(url.searchParams.get('end_ns'), '2000000000');
+  assert.equal(url.searchParams.get('agent_name'), 'claude');
+  assert.deepEqual(response[0].ttft_ms, { p50: 10, p95: 20, p99: 30 });
+  assert.deepEqual(response[0].tps_tokens_per_second, { p50: 40, p95: 50, p99: 60 });
+  assert.deepEqual(response[0].e2e_latency_ms, { p50: 100, p95: 200, p99: 300 });
+  assert.equal(response[0].tpot_ms_per_token, null);
+});
+
+test('fetchLatencyMetrics omits agent_name when no filter is provided', async () => {
+  let requestedUrl = null;
+  global.fetch = async (url) => {
+    requestedUrl = String(url);
+    return new Response('[]', { status: 200 });
+  };
+
+  await fetchLatencyMetrics(3_000_000_000, 4_000_000_000);
+  const url = new URL(requestedUrl);
+
+  assert.equal(url.searchParams.get('start_ns'), '3000000000');
+  assert.equal(url.searchParams.get('end_ns'), '4000000000');
+  assert.equal(url.searchParams.has('agent_name'), false);
+});
+
 test('fetchSecurityCase accepts a valid system-audit detail response', async () => {
   global.fetch = async () => new Response(JSON.stringify({
     state: 'ok',