Building Agentic Workflows

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?

• Generative AI in Contact Centers

• The Role of Data Science in Education

🧠 Building Agentic Workflows

Agentic AI workflows involve building multi-agent systems where a main agent delegates tasks to smaller agents with specific access to tools or data endpoints. These systems are complex, requiring domain expertise, quality data, and appropriate setup. This article demonstrates a tech research workflow using Flowise, which is built upon LangGraph. The workflow answers tech-related questions using data from the last six months.

Agentic AI

Agentic AI refers to multi-agent systems, allowing a large language model (LLM) to engage in System 2 thinking (deliberate and logical), as described by Daniel Kahneman. Using these agents enables more methodical problem-solving by allowing collaboration, planning, and tool access. While useful, these workflows require skill and careful instruction, especially when working with different LLMs and avoiding loop issues.

Economics of Agents

Running agentic workflows can be costly due to the multiple API calls to LLMs like GPT-4. To reduce expenses, smaller models can be used for simple tasks, while more advanced LLMs handle complex reasoning. Optimizing agents' usage and monitoring performance is critical to ensuring the workflow remains cost-effective.

Building Blocks

The workflow is built using Flowise's visual interface, which simplifies the development process for non-coders. Flowise uses a supervisor-worker model, where the supervisor delegates tasks to workers (agents), each performing a specific function with connected tools.

Data Pipeline & Custom Tools

A robust data pipeline is essential for generating valuable responses. The tech research agent uses custom API endpoints to retrieve, analyze, and summarize tech trends. Users can import custom tools in Flowise to enhance functionality.

Debugging & Monitoring

Agent debugging can be challenging, especially in a framework like Flowise, which is still in Beta. Tools like LangFuse help monitor agent performance, API calls, and trace workflow steps to ensure efficiency and proper error handling.

Building agentic AI systems requires patience, careful planning, and a good understanding of how to instruct LLMs through natural language, but it’s a valuable skill for advanced AI tasks.

Top 5 AI-Powered UX Design Tools

1. Uizard

   • Features: Instant wireframing, smart element recognition, code-free design handoff.

   • Limitations: Limited customization, occasional slowness, reliance on templates.

   • Pricing: Free, Pro: $12/month, Premium: $49/month.

2. Attention Insight

   • Features: Instant heatmaps, predictive engagement analysis, platform neutral.

   • Limitations: Confusing dashboard, slow loading, limited customization.

   • Pricing: Solo: €19/month, Teams: €399/month.

3. Khroma

   • Features: AI-powered color suggestions, adaptive color accessibility.

   • Limitations: Limited color options, clumsy interface, exporting hassles.

   • Pricing: Free (beta version).

4. Jasper

   • Features: Smart user flow analysis, instant heatmaps, feedback polls.

   • Limitations: Glitchy responsiveness, outdated templates, limited customization.

   • Pricing: Creator: $39/month, Teams: $99/month.

5. Adobe Firefly

   • Features: Easy prototyping, interactive components, smart design versioning.

   • Limitations: Limited customization, frequent glitches, confusing interface.

   • Pricing: Individual: $4.99/month, Business: $37.99/month.