Configure OpenClaw Thinking Levels: Optimize AI Reasoning

Ever feel like your AI is either rushing through a complex task or overthinking a simple question? It is frustrating when you cannot control the “brainpower” behind a response without rewriting your entire configuration.

You can now use simple directives to tell your agents exactly how much effort to put into their reasoning. Whether you need a quick answer or a deep dive into complex logic, these commands give you direct control over the process.

What it does

You can use an inline directive in any message you send: /t <level>, /think:<level>, or /thinking <level>.

These are the levels and their aliases you can use:

• off | minimal | low | medium | high | xhigh | adaptive
• minimal → “think”
• low → “think hard”
• medium → “think harder”
• high → “ultrathink” (max budget)
• xhigh → “ultrathink+” (GPT-5.2 + Codex models only)
• adaptive → provider-managed adaptive reasoning budget (supported for Anthropic Claude 4.6 model family)

For flexibility, x-high, x_high, extra-high, extra high, and extra_high all map to xhigh. If you want the best possible results, highest and max map to high.

Here are some specific provider notes:

• Anthropic Claude 4.6 models default to adaptive when you do not set an explicit level.
• Z.AI (zai/*) only supports binary thinking (on/off). Any level that is not off is treated as on and mapped to low.
• Moonshot (moonshot/*) maps /think off to thinking: { type: "disabled" }. Any other level maps to thinking: { type: "enabled" }. When thinking is active, Moonshot only accepts tool_choice as auto|none. OpenClaw handles this by normalizing incompatible values to auto.
• General fallback logic applies to other providers.

Resolution order

When you send a message, the system looks for the thinking level in this specific order:

1. Inline directive on the message (this only affects that single message).
2. Session override (set by sending a message that is only a directive).
3. Per-agent default (agents.list[].thinkingDefault in config).
4. Global default (agents.defaults.thinkingDefault in config).
5. Fallback: adaptive for Anthropic Claude 4.6 models, low for other reasoning-capable models, and off for everything else.

Setting a session default

If you want a specific level to stick for your entire conversation, send a message that contains only the directive. Whitespace is allowed. For example:

• /think:medium