Agentic Workflows

Executive Summary

Moving from "Chatbots" to "Autonomous Agents".

Concept	Definition	Key Pattern
Tool Use	LLM calling external APIs (Search, Calculator)	Function Calling / JSON mode
Reasoning	Thinking before acting	Chain-of-Thought (CoT)
Agentic Loop	Self-correction and planning	ReAct (Reason + Act)
Multi-Agent	Orchestrating specialized models	Hierarchical / Peer architecture

---

1. Core Logic Patterns

Chain-of-Thought (CoT)

Encouraging the model to output its intermediate reasoning steps.

• Few-Shot CoT: Providing examples with reasoning.

• Zero-Shot CoT: Simply adding "Let's think step by step" to the prompt.

The ReAct Pattern

A loop of Thought $\rightarrow$ Action $\rightarrow$ Observation.

1. Thought: "I need to find the current weather in Paris."

2. Action: get_weather(city="Paris")

3. Observation: "15C and Cloudy."

4. Conclusion: "The weather in Paris is 15C."

---

2. Compound AI Systems

Instead of one massive model, use multiple components.

• Routing: A small model decides which specialized model to use (e.g., Coding agent vs. Summarization agent).

• Consensus: Multiple agents vote on an answer to reduce hallucinations.

---

3. Cognitive Architectures

• Memory: Giving agents a "working memory" (Short-term context) and "long-term memory" (Vector DB retrieval).

• Planning: Breaking a complex user goal into sub-tasks (e.g., Task Decomposition).

---

Interview Questions

1. "What is an 'Agent' in the context of LLMs?"

An agent is a system where the LLM is the "brain" that uses a planning loop and external tools (APIs, code execution) to accomplish a non-trivial goal. Unlike a simple RAG, an agent can decide how to use the information it retrieves.

2. "How do you prevent an Agent from getting stuck in an infinite loop?"

1. Set a Max Iteration limit ($N=10$). 2. Implement a "Final Answer" override if the model repeats the same thought pattern 3 times. 3. Use an external "Monitor" model to evaluate the loop.

3. "Explain Function Calling."

It's a structured way for the model to output its intent to use a tool. Instead of prose, the model outputs a valid JSON object matching a predefined schema which the developer's code then executes.

---

Agentic Pseudo-code

while not finished:
    thought = model.generate(f"Goal: {goal}\nHistory: {history}\nNext Step?")
    if "FINAL ANSWER" in thought: break
    tool, params = extract_action(thought)
    observation = execute(tool, params)
    history.append(observation)