Building AI Agents

Welcome to learning edition of the Data Pragmatist, your dose of all things data science and AI.

📖 Estimated Reading Time: 5 minutes. Missed our previous editions?

• Consulting Experience and Its Relevance to Data Science

• Understanding Data Contracts

🧠 Building AI Agents

Developing AI agents has been a thrilling process, from the excitement of a working prototype to the frustration of early failures in real-world applications. Achieving stability and performance across various data sources has been a major milestone, though this is still an evolving technology.

What Are AI Agents?

AI agents are autonomous systems built using models like GPTs, Claude, or Cohere’s Command R+. These agents perform tasks by making "tool calls," such as retrieving data or running computations, and iterating through the process to complete a given objective. Unlike scripted systems, AI agents choose their next actions autonomously based on the situation.

Key Lessons Learned

1. Focus on Reasoning Over Knowledge
   Agents should be designed to reason rather than merely possess knowledge. For example, in complex SQL queries, success often comes after multiple failures. Providing the agent with context and allowing it to "think" and adapt ensures that errors are resolved efficiently, ultimately leading to task completion.

2. Importance of Agent-Computer Interface (ACI)
   The ACI, or how the agent interacts with tools, is critical. Small changes in the input/output formats can have a big impact. For example, switching from markdown to JSON format greatly improved performance with certain models. Iterating on ACI can drastically enhance the agent’s ability to process tasks effectively.

3. Model Limitations and Decision Making
   The performance of an agent is limited by the underlying model. For instance, GPT-4’s reasoning abilities outperform GPT-3.5, which may rush into actions prematurely. A better model leads to more accurate decisions, especially in complex workflows involving multiple tool calls.

4. The Drawbacks of Fine-tuning
   Fine-tuning models to improve task-specific performance can backfire by reducing the agent’s ability to reason. It teaches the model to follow patterns, limiting adaptability. However, fine-tuning may still be useful in specific areas, like improving SQL queries within the agent's tasks.

5. Avoid Using Abstractions for Production
   In production environments, it’s better to avoid third-party libraries like LangChain or LlamaIndex. Direct control over input and output calls is essential for debugging, scaling, and ensuring the agent functions as expected.

6. Non-AI Components are Critical
   Beyond the agent’s intelligence, several non-AI elements are vital for success. These include security protocols, data connectors for integrating with external systems, and user interfaces that allow users to follow along and audit the agent’s decisions. Long-term memory also plays a role in improving performance over time.

7. Model Improvements Will Continue
   AI models will keep improving, and agents need to stay adaptable to integrate new advancements. For instance, transitioning to a more powerful model like GPT-4 within 15 minutes of its release ensures competitive edge and better performance.

Additional Tips for Agent Development

• Start Simple: Use tools like pgvector in Postgres for vector similarity search unless a more complex vector database is necessary.

• Don’t Over-Optimize Early: Cost optimization too early can limit scalability and innovation.

• Evaluate Agents Effectively: Establish a framework for evaluating agent performance on both individual tool calls and overall task completion. Streaming tokens can also improve user experience by reducing latency.

AI agents have the potential to transform knowledge work, and building robust, adaptable systems is key to unlocking their full potential.

Top 5 AI Tools for Marketing

1. Jasper AI (for copywriting)

   • Renowned for creating copy across a range of tones and styles.

   • Previously known as Jarvis, inspired by Iron Man’s assistant.

   • Ideal for generating early drafts for human polishing.

   • Highly rated with over 5,000 five-star reviews.

2. Notion AI (for productivity)

   • Integrated into the Notion productivity platform.

   • Helps with tasks like writing, brainstorming, and filling out tables.

   • Data privacy focused, with GDPR compliance and encrypted data.

3. Surfer SEO (for content writing)

   • Optimizes content for search engine ranking.

   • Provides real-time insights on keyword density, readability, and more.

   • Works with various tools like Jasper, WordPress, and Google Docs.

4. Lexica Art (for blog thumbnails)

   • AI-powered image generator for creating realistic visuals.

   • Suitable for marketing content and social media posts.

   • Ideal for creating on-brand blog thumbnails.

5. Content at Scale (for generating SEO blog posts)

   • Specializes in generating SEO-optimized blog content.

   • Produces content that passes AI detectors with a high "human-written" score.

   • Still in development with occasional UI bugs