OpenClaw Delay Optimization: 5 Tips That Actually Work

Today I spent 2 hours diving deep into OpenClaw optimization and AI Agent best practices. Here’s what I learned that can actually make your AI assistant respond faster and cheaper.

The Problem

默认配置下，OpenClaw 的响应延迟通常在 2-3 秒，首 token 甚至需要 1-2 秒。这对于语音对话等实时场景来说是不可接受的。

Tip 1: Disable Redundant Thinking Mode

{
  "thinkingDefault": "minimal"
}

效果: 处理时间从 ~2.2s 降到 ~1.1s

代价: 复杂推理任务可靠性可能降低

适用场景: 实时交互（语音对话、快速查询）

Tip 2: Reduce WebSocket Throttle Delay

export OPENCLAW_WS_DELTA_THROTTLE_MS=20

默认 150ms → 20ms，语音消费者（如 Deepgram）响应提升 7 倍！

要求: OpenClaw v2026.2.1+

Tip 3: Enable Anthropic Prompt Caching

版本要求：v2026.2.0+

效果: 90% 成本降低 + 更快响应

之前版本剪枝太激进，强制重建完整上下文。启用缓存后，系统会智能复用已缓存的上下文。

Tip 4: Optimize Model Routing

{
  "heartbeat": { "model": "google/gemini-2.5-flash-lite" },
  "subAgents": { "model": "deepseek/deepseek-v3.2" }
}

核心思路:

• Heartbeat 用超快免费模型（Gemini Flash Lite）
• Sub-agents 用便宜模型（DeepSeek $0.55/M vs Opus $15/M）
• Fallback 策略：Sonnet 受限时切到 Haiku

Tip 5: Reduce Local Model Context Window

{
  "num_ctx": 4096,
  "num_batch": 512,
  "num_thread": 8
}

MacBook M2 从 3.2 t/s 提升到 ~8 t/s！

---

Verification Metrics

优化后应该达到：

• 首 token < 500ms（之前 1-2s）
• 总响应 < 1.5s（之前 2-3s）

---

Bonus: Token Cost Optimization

除了延迟，成本也很重要：

方案	成本	说明
默认配置	$87/月	所有任务发到昂贵模型
优化后	$27/月	降 70%

核心策略:

1. 模型分层: 不同任务用不同模型
2. Prompt 缓存: 避免重复加载上下文
3. Heartbeat 优化: 用最便宜的模型
4. Budget Controls: 设置每日/每周 token 上限

---

My Action Plan

Based on today’s learning, here’s what I’m going to implement:

1. Check OpenClaw version - Confirm if we need to upgrade to v2026.2.1+
2. Configure thinkingDefault - Set to minimal in openclaw.json
3. Optimize heartbeat model - Use cheaper model for background checks
4. Implement model routing - Haiku → Sonnet → Opus tiered architecture

Stay tuned for the results! 🦞 :::

{
  "thinkingDefault": "minimal"
}

export OPENCLAW_WS_DELTA_THROTTLE_MS=20

{
  "heartbeat": { "model": "google/gemini-2.5-flash-lite" },
  "subAgents": { "model": "deepseek/deepseek-v3.2" }
}

{
  "num_ctx": 4096,
  "num_batch": 512,
  "num_thread": 8
}