Jahnaviio/ai-inbox-env

SmartInbox RL Environment

Unlike traditional spam filters, this environment focuses on sequential decision-making under constraints, making it suitable for training intelligent agents rather than static classifiers.

Overview

This is an OpenEnv-based reinforcement learning environment that simulates real-world email management.

Problem

Managing emails is not just classification — it involves prioritization, decision-making, and trade-offs.

Agents must:

• —identify spam
• —prioritize important emails
• —decide when to respond
• —avoid wasting time on irrelevant messages

Environment Design

This environment models a dynamic inbox where:

• —multiple emails are processed sequentially
• —each action has a cost (time/effort)
• —wrong decisions have penalties

Action Space

• —movetospam
• —respond
• —ignore

Observation Space

Each step provides:

• —subject
• —body
• —sender

Reward Design

Rewards are shaped to simulate real-world trade-offs:

• —Correct classification: +0.7
• —Incorrect decision: -0.4
• —Responding to spam: -0.3
• —Ignoring important email: -0.6
• —Reply cost: -0.1

All rewards are normalized to (0,1) for stable evaluation.

Tasks

• —Easy → clear signals
• —Medium → ambiguous emails
• —Hard → phishing + mixed intent emails

Real-World Impact

This environment can be used to train AI agents for:

• —email assistants
• —productivity tools
• —enterprise inbox automation

Advanced Agent Challenge

This environment is intentionally designed to challenge modern LLM agents:

• —Requires contextual reasoning (not keyword matching)
• —Penalizes unnecessary actions (cost-aware decisions)
• —Includes adversarial inputs (phishing, misleading urgency)
• —Tests long-term reward optimization across multiple steps

Baseline agents achieve moderate scores, but optimal performance requires strategic decision-making.

This makes the environment suitable for benchmarking advanced AI agents beyond simple classification tasks.