Skip to main content

Agent Fundamentals

Agents extend LLMs from simple text generators to autonomous systems capable of reasoning, using tools, and acting on the world.

Core Architectures​

1. ReAct (Reason + Act)​

The ReAct framework interleaves reasoning traces with actions.

  • Thought: The agent thinks about what to do next.
  • Action: The agent decides to use a tool.
  • Observation: The tool returns a result. (This cycle repeats until completion).

2. Plan-and-Execute​

Instead of deciding step-by-step, the agent creates a full plan upfront, executes it sequentially, and adjusts only if a step fails. Great for long-horizon tasks.

3. Reflexion & LATS​

Reflexion adds a self-reflection step where the agent critiques its own past actions. Language Agent Tree Search (LATS) combines MCTS (Monte Carlo Tree Search) with LLMs to explore multiple reasoning paths before acting.

Implementation Example​

python
# Simple ReAct Loop Concept
def react_agent(prompt):
    while True:
        thought, action = llm.think_and_act(prompt)
        if action == "DONE":
            break
        observation = execute_tool(action)
        prompt += f"\nObservation: {observation}"