Email Triage Agent

An LLM-powered agent that reads incoming emails, classifies them by urgency and category, and generates a prioritized action plan with recommended next steps.

Problem

A data scientist receives dozens of emails daily: production alerts, client requests, project updates, newsletters. Manually prioritizing these wastes time and risks missing critical issues. This agent automates the triage process.

How It Works

Incoming Emails → [Classification] → [Prioritization] → Action Plan
                   urgency             sort by           recommended
                   category            urgency           next steps

Classification

Each email is classified on two dimensions:

• Urgency: critical / high / medium / low / ignore
• Category: production_issue, client_request, security, project, hr_admin, billing, newsletter

Output

A prioritized action plan with:

• Summary of each email
• Recommended action (specific, not vague)
• Estimated time to handle
• Whether a reply is needed

Example Output

[!!!] CRITICAL   | Urgent: Model predictions seem off
      Category: production_issue
      Summary: Credit scoring model returning abnormally high default probabilities
      Action: Check model deployment logs and compare feature distributions
      Time: 2hr [REPLY NEEDED]

[ ! ] HIGH       | Feature request: explainability report
      Category: client_request
      Summary: Client needs SHAP values for regulatory compliance
      Action: Scope SHAP integration into API response, estimate timeline
      Time: 30min [REPLY NEEDED]

[ - ] MEDIUM     | New project: churn prediction model
      Category: project
      Action: Accept meeting invite, prepare data availability overview
      Time: 15min

Usage

# Rule-based mode (no API key needed)
python agent.py

# With OpenAI
export OPENAI_API_KEY=your_key
python agent.py --provider openai

# With Anthropic
export ANTHROPIC_API_KEY=your_key
python agent.py --provider anthropic

# Custom email file
python agent.py --emails path/to/emails.json --output results.json

Architecture

agent.py
├── Data models (Urgency, Category, TriageResult)
├── Prompt engineering (SYSTEM_PROMPT, build_triage_prompt)
├── LLM providers (OpenAI, Anthropic, local fallback)
└── EmailTriageAgent
    ├── triage_email()      → classify single email
    ├── triage_batch()      → classify + sort by urgency
    └── generate_action_plan() → formatted output

Design Decisions

1. Provider-agnostic — Swappable LLM backends (OpenAI, Anthropic, or rule-based fallback). The agent logic doesn't depend on any specific API.

2. Structured output — Forces JSON responses with a predefined schema. This makes the output parseable and actionable, not just free-text summaries.

3. Local fallback — Keyword-based classification works without any API key. Useful for testing the pipeline and demonstrates that the architecture matters more than the model.

Tech Stack

Python, OpenAI API, Anthropic API