Reasoning Effort
How hard an agent thinks before it answers. One scale across every provider, narrowed to what each model actually accepts.
mur agent effort researcher # show the current setting
mur agent effort researcher xhigh # a single-purpose specialist earns it
mur agent effort worker low # a fan-out worker does not
mur agent effort researcher --clear # back to the provider default
Inside a chat, /effort changes it for the conversation without a restart:
/effort # levels this model accepts, current value, and where it came from
/effort high # this session, effective next turn
/effort high --save # this session and the profile
The MUR Hub has the same control in an agent's Behavior tab, where it writes the profile — a settings pane is a statement about the agent, a slash command is about one conversation.
The scale
low · medium · high · xhigh · max, cheapest first.
Leaving it unset is not "no effort" — the provider default is high, so every unset agent is already paying for high effort. The useful direction for mechanical work is down.
Not every model takes every level
The levels a model accepts are a property of that model, not a fixed list, so /effort and the Hub only ever offer what will work:
| Model family | What it takes |
|---|---|
| Claude (Opus 4.7+, Sonnet 5) | the whole scale |
| Claude (Opus 4.5 / 4.6, Sonnet 4.6) | everything except xhigh — that step arrived in 4.7 |
| GPT-5, o1 / o3 / o4 | low · medium · high |
| DeepSeek V4 | low · high · max — there is no middle step |
| Grok 4.6+ | low · medium · high · xhigh |
| Grok 4.3 / 4.5 | low · medium · high |
| Gemini 3 | low · medium · high |
| Gemini 2.5 | the same levels, converted to a token budget |
| Qwen 3, GLM | a thinking switch, not a dial — two positions |
| Magistral | always reasons, and rejects the parameter |
| Everything else | no reasoning control |
A level a model has no step for is degraded to the nearest one it does have, never sent and never an error. If you set medium on DeepSeek V4, you get low and MUR tells you so rather than leaving you to believe a setting that never applied.
Where the setting lives
effort: on the agent's profile.yaml, because it describes the agent's job. A /effort change without --save is session-only and does not touch the file.
Individual mechanical calls inside the runtime cap themselves lower regardless of what is set here.
Changing the model can change the level
The stored level and the level in force are separate. Set an agent to xhigh on Claude, switch it to DeepSeek V4, and the agent runs at high — but the profile still holds xhigh, so switching back restores it. The Hub says which is which instead of showing the narrowed value as if it were the setting.