Agent Conversation Psychology Report

Problem It Addresses

Multi-turn agent conversations often break down in subtle ways: a user repeats the same request, a long answer loses the thread, or the agent rushes to retry after a failure. These collaboration frictions are easy to feel but hard to diagnose. Agent Psychology Report treats a single session as the unit of review, turning 'something felt off' into observable behaviors, theoretical explanations, and actionable suggestions for both sides.

How It Works

• Trigger modes: It can be proposed during a conversation, explicitly requested with phrases such as generate a review report, or loaded through keywords like Agent Psychology Report and conversation diagnosis.
• Scope confirmation: It first checks whether to review the recent exchange or the full session, avoiding overgeneralized conclusions.
• Context analysis: Instead of relying on memory_recall or historical retrieval, it reads the current context directly, extracts 2–4 symptom keywords, and maps them to collaboration issues.
• Theory mapping: It connects observations to entries such as working_memory, cognitive_load, emotional_contagion, and attribution_bias, then proposes AI-side fixes and optional user-side adjustments.
• Feedback loop: Validated conventions can be stored as neutral memory notes, such as 'lead long answers with three points,' so the review becomes a reusable collaboration rule.

Boundaries

It is best used inside the current session for diagnosis and collaboration tuning, not as a background job that pulls history. Reports should describe observable behavior rather than label users with diagnostic terms. Sensitive or emotional content should be abstracted, and incomplete context should be flagged with confidence limits.