As the focus shifts to AI Agents/Agentic Workflow, it is expected that agents will increasingly be implemented to increase worker productivity and re-imagining business processes. Large Language Models (LLMs) are being enhanced with reasoning/thinking and planning abilities as well as the ability to take actions. There is so much information that it can at times become somewhat overwhelming. In this blog, we explain what AI agents are, popular use cases, and how to develop AI agents.
What is an AI Agent?
At its simplest, an AI agent can be thought of as an Augmented LLM that has access to retrieval, memory, tools.
Agents are emerging as LLMs improve in core capabilities (understanding, reasoning/thinking, tool use, and error recovery). They have the following features:
- Work based on user commands or interactive discussions.
- Operate independently with potential human feedback.
- break down complex tasks into sub-tasks
- Iterative cycle of “thought, action, and observation”
- Rely on “ground truth” from the environment.
- Use stopping conditions for control.
- Implemented as LLMs using tools in a loop.
- Require well-designed and documented toolsets
Agent Workflow:
The user starts the interaction and the flow is as follows:
- Human input (command or discussion).
- Independent planning and operation by the Agent.
- Environmental feedback (ground truth).
- Human feedback (optional checkpoints, blockers).
- Task completion or stopping conditions.
Where can you use AI agents?
AI agents are in their infancy right now and they are expected to impact almost every line of business including Finance, HR, Supply Chain, Manufacturing, and Coding. As examples, Uber and Replit are using AI Coding agents for software development and testing. AI agents are being used for Research, Knowledge Management, and Customer Service by others.
How do you Build an AI Agent?
AI agents can be developed by augmenting LLMs via with access to tools and functions that enable them to take actions. These tools could be like web search tools or access to APIs. If you prefer, you could also use Agent building frameworks from LangChain and others. At Unvired, we have used LangGraph from LangChain to build AI agents for Agentic Retrieval Augmented Generation (Agentic RAG) and other use cases.
LangGraph provides a library for creating agent and multi-agent workflows. Some of the key features of LangGraph are:
- Cycles and Branching: Implement loops and conditionals
- Persistence: Auto-save state after each step, Pause and resume execution
- Human-in-the-Loop: Interrupt for approval or editing of agent actions
- Streaming Support: Stream outputs as they’re produced (including token streaming)
- LangChain Integration: Seamless integration with LangChain and LangSmith
LangGraph also has templates that provide easy entry points for building and deploying sophisticated agentic applications. These address common use cases with configurable templates for Python and JavaScript. Templates are structured for easy debugging in LangGraph Studio and one-click deployment to LangGraph Cloud. Configurable templates allow choosing providers (language models, vector stores, tools) without vendor lock-in.
We hope that this helps you thinking about your own AI agents journey. The earlier we all start re-imagining business processes powered by AI Agents, the greater the chance that we grab the first mover advantage and create disruption. Failing which, we face the risk of being disrupted. What would you rather do-disrupt or get disrupted?
As models like OpenAI o1 get even better with their reasoning, AI Agents will get more intelligent and more compelling. And Agent Computer Use/Interface from Claude and others is already available in pilot mode. In our next blog, we will examine how o1 or o1-mini can be used to plan and execute tasks to deliver even more intelligent AI agents as well as the potential that Agent Computer Interface has to change how we work.
