AI Agents

Agents differ from workflows in their autonomy. While a workflow follows a pre-defined path, an agent decides its own path.

The Loop

An agent typically operates in a loop:

1. Observe: Look at the user goal and current state.
2. Think/Reason: Decide what to do next (e.g., "I need to search Google").
3. Act: Execute a tool (run the search).
4. Evaluate: Look at the result. Is the goal met? If not, loop back to step 1.

ReAct Pattern

Reasoning + Acting. The agent explicitly writes down its thought process before taking an action.

Thought: The user wants the weather in Tokyo. I should use the weather tool.Action: get_weather("Tokyo")Observation: It is 15°C and raining.Thought: I have the answer.Final Answer: It's currently raining and 15°C in Tokyo.

Multi-Agent Systems

Using multiple specialized agents that collaborate.

• Manager Agent: Breaks down the plan and assigns tasks.
• Coder Agent: Writes code.
• Reviewer Agent: Checks code.

Challenges with Agents

• Loops: Getting stuck repeating the same action.
• Cost: Many steps = many tokens.
• Unpredictability: Harder to debug than linear workflows.