Does lower reasoning effort mean worse results?

For simple tasks, no. Low reasoning is fine for file operations, git commands, and straightforward queries. Save high reasoning for architecture decisions and complex debugging.

Can Hermes automatically detect task complexity?

Not yet built-in, but the proposed 'auto' mode would let the agent assess complexity before choosing reasoning level. Community testing is ongoing.

How much token savings are we talking about?

Depends on usage pattern. Heavy users report 20-40% token reduction when manually switching reasoning levels for simple vs complex tasks.

Adaptive Reasoning Effort — Let Hermes Control Its Own Thinking

Not every task needs maximum reasoning effort. A quick file rename shouldn't burn the same tokens as architecting a distributed system.

A community member proposed a plugin that lets Hermes control its own reasoning level on the fly — and the discussion reveals how this could work.

The Problem

Current Hermes config sets agent.reasoning_effort globally. If you have multiple sessions running:

One coding session needs reasoning_effort=high
Another session doing simple file ops needs reasoning_effort=low
They're forced to share the config level

The Proposed Solution

Persist reasoning effort at the session level, not config level. Each session tracks its own reasoning setting.

The agent could then have a tool to adjust it:

/set-reasoning high   # Complex architecture work
/set-reasoning low    # Simple file operations
/set-reasoning auto   # Let Hermes decide per-task

Implementation Approaches

Option 1: Plugin with Agent Access

The plugin would need to:

Access the live AIAgent instance from the tool handler, OR
Mutate outbound request reasoning config via a hook

This requires core changes to expose the agent instance to plugins.

Option 2: Skill with Config Commands

A skill that tells the agent to use hermes config set agent.reasoning_effort <level>. Simpler but affects all sessions.

Option 3: Session-Level Persistence (PR Open)

The community member has PR #7345 open that would persist reasoning level per-session. This is the cleanest solution.

Why This Matters

Token costs add up. If Hermes could recognize:

"This is a simple git status, use low reasoning"
"This is a complex refactor, use high reasoning"

...you'd save significant tokens over a month of usage without sacrificing quality on complex work.

Current Status

The Nous team is evaluating the PR. From Discord, the maintainer noted: "I have an extremely high bar at this point to adding any tools" due to complaints about tool context use.

The skill approach might ship first as a community plugin while the core feature is debated.

Try It Yourself

If you want to experiment with manual reasoning switching:

# Before a complex task
hermes config set agent.reasoning_effort high

# After, for simple tasks
hermes config set agent.reasoning_effort low

Not automatic, but demonstrates the token savings. Watch the community-projects-showcase channel for when the automated plugin ships.

Adaptive Reasoning Effort: Token-Smart Thinking