Agentic Systems
Explore agentic systems in AI engineering to understand how autonomous agents differ from chatbots by using tools, maintaining state, and planning multi-step tasks. Learn about the ReAct reasoning framework, tool integration, protocols like MCP and A2A, multi-agent architectures, common failure modes, and how to implement a minimal ReAct agent loop to prepare for AI-focused interviews.
We'll cover the following...
- What is an AI agent and what distinguishes it from a chatbot?
- What is ReAct and why is it the foundational agentic technique?
- How does tool use work in an agent?
- What is the Model Context Protocol and why does it matter?
- What is A2A and how does it complement MCP?
- What are multi-agent systems and what are their failure modes?
- How would you implement a minimal ReAct agent loop?
- What’s next?
If there is one topic that has shifted the conversation in AI engineering most dramatically from 2024 to 2026, it is agents. The industry has moved from single-turn Q&A to systems where models plan, use tools, coordinate with other agents, and execute multi-step tasks with real-world consequences. Interviews at AI-forward companies increasingly include agent architecture questions, and “I’ve used LangChain” is not a sufficient answer. You need to understand the underlying patterns, the failure modes, and the emerging standards that are shaping how agents interoperate.
An agent is not a smarter chatbot. The fundamental difference is autonomy and side effects. A chatbot generates text and stops. An agent can read and write files, call APIs, send emails, execute code, and make purchases. The stakes of getting the behavior wrong are proportionally higher, and the engineering discipline required is proportionally more rigorous.
What is an AI agent and what distinguishes it from a chatbot?
A chatbot operates in a single-turn or multi-turn conversation loop: user sends message, model generates response, done. The model has no persistent state, takes no external actions, and has no ability to affect the world beyond the text it produces.
An agent operates in a perceive-decide-act-observe loop. It receives an initial task or goal, reasons about what actions to take, executes those actions using available tools, observes the results, and iterates until the task is complete or it determines it cannot proceed. Crucially, the loop can run for many steps without human intervention.
An agent has three defining properties:
Tools: Mechanisms to take actions in the world.
State: Memory of what has happened so far, either in context or in external storage.
Objective: A goal that drives multi-step planning rather than a single response.
The most common tools are web search, code execution, file read/write, database queries, API calls, and calls to other LLMs. The model does not execute these tools; it outputs a structured request, and the host application executes them. The result is returned to the model as an observation, and the loop continues.
What is ReAct and why is it the foundational agentic technique?
ReAct (Reasoning ...