Chatbots and Agency
Chatbots use conversational Natural Language Processing techniques to understand user queries (inputs | prompts) and automate responses. The chatbot itself is the medium through which the user interacts, while “agency” is the chatbot system's technological framework that lets it connect with other tools, plan, execute actions and learn from past actions.
Non-agentic chatbots don't have access to tools, nor do they have memory to store learnings or actions. They require input from their users. They're limited to producing responses to common prompts aligned with user expectations, but perform poorly on specific questions. Since they have no memory, they can't learn from their mistakes when their responses fall short.
By contrast, agentic chatbots learn over time to adapt to user expectations and create personalized experiences. They can achieve goals by completing complex tasks and subtasks without human intervention and consider different approaches, which can also be corrected or updated as needed. Unlike non-agentic chatbots, these chatbots evaluate the tools available to them to fill information gaps.
Reasoning Approaches
There's no single standard architecture for building AI Agents. There are several approaches for solving problems that require multiple steps or stages.
ReAct (Reasoning and Action)
In this approach, the agent is configured to “think” and “plan” after each action it takes and, with each response tool it has available, decide which one to use in the next step. This Think – Act – Observe cycle is used to solve goals step by step and improve with each response.
Through prompt structure, agents can be instructed to reason in a paced way and show each “thought.” The agent's verbal reasoning gives an idea of how its responses are formulated. In this framework, agents continuously update their context with new reasoning. This approach can be understood as a form of Chain-of-Thought prompting.
ReWOO (Reasoning Without Observation)
The ReWOO method, unlike ReAct, removes the dependency on tool outputs for planning actions. Instead, agents plan ahead of time. Redundant tool use is avoided by anticipating which tools will be needed upon receiving the user's initial prompt. This approach is desirable from a human-centered perspective, since the user can confirm the plan before it's executed.
The ReWOO workflow is made up of three modules. In the planning module, the agent anticipates the steps to follow based on the user's prompt. The next stage involves collecting the outputs produced by executing the planned tools. Finally, the agent combines the initial plan with the tool outputs to formulate the final response. This upfront planning can significantly reduce token usage, computational complexity and the consequences of intermediate tool failures.
Up to this point, we've seen how a chatbot stops being just a natural language processor and becomes an agentic system, capable of planning, executing actions and adapting over time. The difference isn't in the conversational interface, but in the internal architecture that enables reasoning, tool use and decision-making.
Paradigms like ReAct and ReWOO show that there's no single way to build AI agents, but rather different approaches for solving problems made up of multiple steps, with different levels of control, efficiency and autonomy.
In the next article, we'll look at the types of AI agents, from the simplest to agents with learning capabilities, to understand how they're classified based on their capabilities and in which contexts each one is appropriate.