PersuasionRL
My project is about Adaptive Multi-Agent Deception: Can AI Learn to Sell to Adversarial Personalities?
Core trick
Sales agent with closed-loop LLM oversight writes executable policy hints from failure analysis between RL episodes.
Tags
Description
My project is about Adaptive Multi-Agent Deception: Can AI Learn to Sell to Adversarial Personalities?
The Challenge: The prospect is adversarial. it has hidden traits (price_sensitivity, relationship_importance, decision_speed) that determine whether sales rep's actions work or backfire. The sales agent must learn to:
* Infer hidden state (is this prospect relationship-driven or transactional?)
* Adapt strategy mid-episode
* Generalize across personalities (learn different policies for 8+ prospect archetypes)
Novel Contribution: Closed-loop LLM oversight that analyzes failure modes and writes executable policy hints, creating a human-AI-RL feedback loop.
AraavNayak/Persuasion-Agent-RL-Environment★ 00 forksPythonCreated Mar 8, 2026 · last push May 26, 2026
README excerpt
# Persuasion Agent RL Environment Auto-playing B2B sales RL environment with expert-in-loop oversight, achieving 10/10 engagement scores through sequential learning. <img width="1512" height="982" alt="Screenshot 2026-03-12 at 12 30 04 AM" src="https://github.com/user-attachments/assets/b5d9b927-d3a8-4fb8-82c0-2565e5430a7d" /> ## 🌟 Key Features ### 1. Auto-Playing Visual Interface - **Sequential action execution** with 1-2 second highlighting - Watch the agent learn in real-time - Actions are highlighted as they're chosen - Episode resets and learns after each deal close ### 2. Fleet AI Oversight Agent - Monitors sales agent behavior for ethics and best practices - Detects spammy patterns, premature discounting, poor timing - Provides real-time alerts and recommendations - Generates episode reports with behavior scores ### 3. Engagement Score Goal (10/10) - **Primary metric**: Get engagement score to 10/10 - Combines warmth, interactions, progress, and Fleet AI behavior score - Clear visual feedback on performance - Learns to optimize for high engagement ### 4. Learning & Iteration - Agent learns from each episode - Resets automatically after deal closes - Iterates to improve engagement scores - Includes Unsloth/TRL training script for Colab ### 5. OpenEnv 0.2.1 Compatible - Built on latest OpenEnv stable release - Ready for HuggingFace Spaces deployment - WebSocket-based real-time communication - Docker deployment included ## 🚀 Quick Start ### Run Locally ```ba