Repository navigation
Harness: [Improvement]: Default instructions should request concise rationale/progress instead of thought process #8553
Description
Summary
The default Harness instructions in both .NET and Python currently ask the model to:
Explain your reasoning and thought process
and also ask it to keep the user informed about its "thought process" between tool calls.
For a provider-neutral agent framework, the default should request concise progress, rationale and results, not hidden/internal reasoning or chain-of-thought.
Why this matters
Framework defaults are inherited by applications unless they explicitly override HarnessInstructions.
The current wording can:
- encourage models to expose private/internal reasoning;
- behave inconsistently across providers that treat reasoning differently;
- cause applications to override the whole Harness instruction block just to avoid this behavior;
- conflate useful user-facing progress with hidden chain-of-thought.
We encountered this while integrating HarnessAgent into a production assistant and replaced the default Harness instructions even though the rest of the defaults were useful.
Suggested direction
For example:
Keep the user informed with concise progress updates when useful.
Explain decisions, actions, results, and important trade-offs at a high level.
Do not expose private chain-of-thought or hidden reasoning.
The exact wording can follow project conventions.
Scope
This appears in both:
- .NET
HarnessAgent.DefaultInstructions - Python
DEFAULT_HARNESS_INSTRUCTIONS
No new abstraction is required.
Code Sample
Language/SDK
Both
Metadata
Metadata
Assignees
Labels
Type
Projects
- Status
Activity
[-]Harness: [Improvement]: Default instructions should request concise rationale/progress instead of thought process[/-][+].NET: Harness: [Improvement]: Default instructions should request concise rationale/progress instead of thought process[/+][-].NET: Harness: [Improvement]: Default instructions should request concise rationale/progress instead of thought process[/-][+]Python: Harness: [Improvement]: Default instructions should request concise rationale/progress instead of thought process[/+][-]Python: Harness: [Improvement]: Default instructions should request concise rationale/progress instead of thought process[/-][+]Harness: [Improvement]: Default instructions should request concise rationale/progress instead of thought process[/+]1aifanatic commentedon Oct 3, 2026
I will update the .NET and Python Harness defaults to request concise progress updates and high-level rationale, and align the data-processing sample instructions with that wording. This will keep the existing tool-use guidance and instruction override behavior.
eavanvalkenburg commentedon Oct 7, 2026
We are marking this issue for vNext this would be a breaking change that would be hard for people to understand and see, and thus breaking it while there is a path to customize these instructions is not the right approach (#9004 was merged on accident, will revert). When we pick this up, it will involve running the different prompts against different models in order to get a good result across the board, but these instructions will never be a one size, fits all.
1aifanatic commentedon Oct 21, 2026
Understood. I'll leave the default-prompt change for vNext and keep #9112 closed. The existing instruction override remains the current customization path; any future default change should follow the team's model-comparison evaluation.