Problem
Some models in Ollama's library (e.g. granite4.2) and other providers now expose a "thinking"/"reasoning" mode that runs before the actual answer. For simple tasks this adds latency and token cost without quality benefit. There's no way to turn it off or tune it from an axe agent today.
Desired Behavior
Let an agent's config (and a CLI override) disable thinking/reasoning or set its effort level, so the same axe agent definition can opt out of thinking on cheap tasks and opt in on hard ones.
Configuration
Something like:
[params]
thinking = "off" # or "low" / "medium" / "high", or just disabled
And/or a CLI override:
axe run my-agent --thinking off
Notes
- Should apply to any provider/model that supports it (Ollama, Anthropic, OpenAI o-series, Gemini, Bedrock Claude 3.7+, etc.) — not just Ollama.
- If a model/provider doesn't support this knob, the setting should be silently ignored (or the agent should fail with a clear message — pick whichever fits axe's style).
- Names and accepted values can be normalized across providers; the user shouldn't have to know provider-specific spellings.
Related
Problem
Some models in Ollama's library (e.g.
granite4.2) and other providers now expose a "thinking"/"reasoning" mode that runs before the actual answer. For simple tasks this adds latency and token cost without quality benefit. There's no way to turn it off or tune it from an axe agent today.Desired Behavior
Let an agent's config (and a CLI override) disable thinking/reasoning or set its effort level, so the same axe agent definition can opt out of thinking on cheap tasks and opt in on hard ones.
Configuration
Something like:
And/or a CLI override:
Notes
Related