Introduction
While a standard LLM call is zero-shot (one input, one output), an agentic workflow involves a loop of reasoning and acting. This allows the system to handle tasks that require multiple steps and course-correction.
Key Components of an Agent
- Planning: The agent breaks down a large goal into smaller, manageable sub-tasks. Techniques like ReAct (Reason + Act) are commonly used here.
- Memory: Agents need to remember what they've done. This includes Short-term memory (the current conversation context) and Long-term memory (often implemented with Vector Databases).
- Tools (Action): The ability to interact with the outside world. This could be searching the web, running code in a sandbox, or querying a database.
- Reflection/Critique: The agent evaluates its own output or the result of a tool call and decides if it needs to try again or take a different approach.
Assignment
- Read the original ReAct paper (or a simplified summary) to understand how combining reasoning traces with actions significantly improves LLM performance.
- Draw a flowchart of an agentic loop (Goal -> Plan -> Action -> Observation -> Critique -> Result).