How to build agents without coding experience

Plus: Microsoft’s hub for cutting-edge agent tech, a breakthrough in humanoid robots, and more

Author

Michael Cunningham
February 24, 2025

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

Agent strategy bullet point #1: subscribe to high-quality publications offering all the latest news on AI agents in the enterprise. I can think of a good place to start.

In today’s issue…

  • Low-code frameworks allow easy agent building

  • Create AI agents with just a description

  • The next 10 years will be about the agent economy

  • Automatically selecting the best agent LLM

…and more

🔍 SPOTLIGHT

Created by the author using Dall-E 3

At first glance, AI agents seem like magical alchemy to many non-coders. But a growing number of low-code tools are making them not only accessible to those outside of software engineering, but easy.

The first true LLM agents, such as AutoGPT and BabyAGI were demoed in the weeks following the release of GPT-4 in March 2023. These systems were inflexible proofs-of-concept with built-in tools and defined workflows, with limited to nonexistent customization options. While users could assign tasks to the agent, such as “write me a report on foreign exchange markets,” changing its functionality in any meaningful way required extensive software engineering knowledge and hours of work.

This began to change as agent frameworks emerged that allowed coders to define their own agents in Python, equip them with custom tools, and arrange them into arbitrary workflows with multiple agents interacting to accomplish tasks. Packages such as LangChain and AutoGen gave programmers the ability to assemble modular components into increasingly sophisticated multi-agent systems. Nevertheless, these packages still required fluency in Python, keeping agents beyond the reach of many of the knowledge workers who would most benefit from their capabilities.

In parallel, however, a new class of agent framework was born: low-code. Low-code tools were not new—their use for workflow automation was a significant trend in the mid-late 2010s—but coupling them with LLMs gave them extraordinary new power. As with the transformation of robotic process automation (RPA), AI agents gave a struggling field the key component it needed to begin changing the face of enterprise work.

In a standard low-code agent application, components of an agentic system, such as LLM calls, databases, web APIs, and more are represented as blocks on a canvas, which can be chained together to create workflows. Rather than convoluted and scattered lines of code, the flow of data through an agent application can easily be visualized step-by-step.

Some of these systems exist as standalone applications, such as n8n, Flowise, Langflow, and Lyzr. In many cases, they allow users to insert custom components that incorporate Python, JavaScript, or other code—thus, while users do not need to be programmers to use them, they have the option to collaborate with programmers to add integrations and functionality.

Others are the low-code implementation of existing code-based frameworks, including AutoGen, CrewAI, and LangGraph. Each of these dominant Python agent frameworks has a corresponding “studio” application that enables workflows to be constructed without code and run directly from the low-code application or exported, allowing non-programmers to build agentic apps and then pass them to software engineers to be put into production.

Finally, enterprise agent providers such as Microsoft, Salesforce, and a host of RPA companies are creating their own agent studio applications, making agentic AI accessible to users of their platforms.

In an interview last year, NVIDIA CEO Jensen Huang made a stir with his bold claim that companies of the future would “employ” thousands of AI agents for every human. As this dream comes closer to reality, it is increasingly clear that the builders of these agents will not be tiny teams of specialized software engineers, but the everyday employees whose work they will be streamlining.

If you find Building AI Agents valuable, forward this email to a friend or colleague!

📰 NEWS

Source: Microsoft

Azure AI Foundry Labs is a new showcase and community for developers and companies to access cutting edge research by Microsoft, including open-source agent software such as multi-agent system Magentic-One and GUI automation tool OmniParser V2.

Robotics startup Figure AI demoed Helix, a revolutionary model for powering humanoid robots that represents a significant step towards taking AI agents out of the virtual world and into the physical one.

🛠️ USEFUL STUFF

Source: AutoAgent

AutoAgent is a zero-code framework that allows users to describe an AI agent they want built using natural language and have it created by another “agent-building agent”.

Graphiti, by YC-backed agent memory startup Zep, is a knowledge graph layer that provides long-term memory for agentic systems, allowing them to recall past data and chats.

This guide shows how to easily create a LangGraph-powered research agent that runs in the cloud on IBM’s watsonx.ai platform

💡 ANALYSIS

Created by the author using Dall-E 3

Agent builders who master the new marketplaces of agentic enterprise software will reap massive rewards, this piece predicts—but only if they survive their ruthless winner-take-all dynamic.

AI agents create a dilemma in enterprise cybersecurity, as their power—and their dangers—both stem from their need to access critical confidential data within an organization.

In a somewhat unusual forum, Scale AI CEO Alexander Wang discusses the present and future promise of AI agents on a longform podcast with comedian Theo Von.

An article by the World Economic Forum arguing that AI agents will have a transformative impact on small- and medium-sized enterprises (SMEs), particularly for their ability to manage complex supply chains.

🧪 RESEARCH

Diagram of LLM selection for compound AI systems | Source: arXiv

AI agents are a kind of compound AI system, in which multiple supporting components are used to enhance an LLM’s capabilities. This paper proposes an automated method for selecting the optimal LLM to maximize agent performance.

MLGym is an effort to create LLM-powered agents that perform research in the AI domain, finding that they can optimize existing models but not yet generate novel scientific hypotheses.

Thanks for reading! Until next time, keep learning and building!

What did you think of today's issue?

Login or Subscribe to participate in polls.

If you have any specific feedback, just reply to this email—we’d love to hear from you

Follow us on X (Twitter), LinkedIn, and Instagram