Agent Mode: Ensure consistent behavior across models

[CarGDev]

Problem

Different LLM models behaved inconsistently during agent execution:

• Some showed plans, others didn't
• Approval flow varied by model
• Output formatting differed

Observed Inconsistencies

Model	Plan Shown	Waited for Approval	Output Format
GPT-5-mini	Yes (multiple)	No	Plain text
Claude Sonnet 4.5	Yes (structured)	No	Markdown
Gemini	Yes	No	Plain text

Expected Behavior

Regardless of which LLM provider is used:

1. Same plan format displayed
2. Same approval flow enforced
3. Same tool calling interface
4. Consistent output formatting

Implementation Suggestions

The agent layer should normalize behavior:

• Wrap model responses in consistent format
• Enforce approval gates at application level (not model level)
• Standardize output through formatters

// Application-level enforcement, not model-dependent
const executeWithApproval = async (plan: Plan): Promise<void> => {
  const approved = await showPlanAndWaitForApproval(plan);
  if (!approved) return;
  // Execute...
};

Priority

🟡 Medium - Affects user experience and predictability

---

Generated from model evaluation test