Agent Architectures

Without explicit reasoning, agents jump to tool calls based on surface patterns — calling the wrong API, misinterpreting results, or skipping necessary steps. Reasoning forces the LLM to articulate its logic, dramatically improving accuracy on complex, multi-hop tasks.

Reasoning is the agent showing its work on a math test — the final answer matters, but the step-by-step logic is what prevents careless errors and makes mistakes debuggable.

Scenario

Question: 'Is our AWS spend trending up or down compared to last quarter?' Agent reasons: need current quarter spend → need last quarter spend → calculate delta → check if increase is across all services or one outlier → then answer with evidence.

Solution

In Agent Foundations, apply Agent Architectures to this scenario: Question: 'Is our AWS spend trending up or down compared to last quarter?' Agent reasons: need current quarter spend → need last quarter spend → calculate delta → check if increase is across all services or one outlier → then answer with evidence. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Agent Architectures (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Agent Architectures happens in the code.

1REASONING_PROMPT = """Solve this task step by step.2
3Before each action, write your reasoning:4Thought: [analyze current state, what you know, what you need]5Action: [tool_name]6Action Input: [parameters]7
8After seeing results:9Observation: [what you learned]10Thought: [updated reasoning]11
12Task: {task}"""13
14messages = [15    {"role": "system", "content": REASONING_PROMPT.format(task=user_task)},16    {"role": "user", "content": user_task},17]

• No reasoning step — agent jumps to wrong tool calls
• Unbounded reasoning consuming entire context window
• Not logging reasoning traces — failures are opaque
• Using expensive reasoning models for simple tool selection
• Free-form reasoning without structured format

Quick recap — the most important points from this module.