Deep Expertise Track · Lesson 1

The Agent Spectrum

5 levels of AI systems — from single LLM call to multi-agent orchestration

The Agent Spectrum: From Single LLM Call to Autonomous Agent

Lesson 1 of the Deep Expertise Track — understanding the 5 levels of agentic systems

What you'll learn
  1. The 5 levels of AI system complexity (from direct call to multi-agent)
  2. Where your existing webapps sit on this spectrum (and why)
  3. The critical distinction between "workflows" and "agents" — the #1 thing hiring managers test for
  4. Why Anthropic says "start simple" and when NOT to use agents

The Big Picture: 5 Levels of AI Systems

Anthropic, the company behind Claude, spent a year working with dozens of teams building LLM agents across industries. Their conclusion? "The most successful implementations weren't using complex frameworks or specialized libraries. Instead, they were building with simple, composable patterns."

Source: Anthropic — Building Effective Agents (Dec 2024)

Before you can build agents, you need to know where agents sit in the broader landscape. Here's the full spectrum — from dumbest to smartest:

┌─────────────────────────────────────────────────────────────────────┐ │ THE 5 LEVELS OF AI SYSTEMS │ │ │ │ LEVEL 1: DIRECT MODEL CALL │ │ ┌──────────┐ │ │ │ "Summarize│ No tools. No loop. No memory. │ │ │ this text"│ Just prompt in, text out. │ │ └────┬─────┘ │ │ ▼ │ │ LEVEL 2: AUGMENTED LLM │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │ LLM │────▶│ Retrieval│ │ Tools │ │ │ │ (brain) │ │ (RAG) │ │ (APIs) │ │ │ └──────────┘ └──────────┘ └──────────┘ │ │ LLM can search docs, call APIs, but YOU pre-script what it calls. │ │ ▼ │ │ LEVEL 3: WORKFLOW (predefined steps) │ │ ┌─────────┐ ┌─────────┐ ┌─────────┐ │ │ │ Step 1 │───▶│ Step 2 │───▶│ Step 3 │ │ │ │ (LLM) │ │ (LLM) │ │ (LLM) │ │ │ └─────────┘ └─────────┘ └─────────┘ │ │ YOU define the order. LLM executes each step. No deviation. │ │ ▼ │ │ LEVEL 4: AGENT (LLM decides what to do) │ │ ┌──────────────────────────────────────────┐ │ │ │ AGENT LOOP │ │ │ │ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ │ │ │ │ │Think│─▶│ Act │─▶│Obser│─▶│Think│ │ │ │ │ └─────┘ └─────┘ └─────┘ └─────┘ │ │ │ │ └────────────────────┘ │ │ │ └──────────────────────────────────────────┘ │ │ LLM decides which tool to call, in what order, and when done. │ │ ▼ │ │ LEVEL 5: MULTI-AGENT ORCHESTRATION │ │ ┌─────────┐ │ │ │MANAGER │───┐ │ │ │ AGENT │ ├─▶ ┌────────┐ │ │ └────┬────┘ ├─▶ │ Agent A│ (specialist) │ │ │ ├─▶ │ Agent B│ (specialist) │ │ │ └─▶ │ Agent C│ (specialist) │ │ └────────────└────────┘ │ │ Multiple agents coordinate. Manager delegates, workers execute. │ └─────────────────────────────────────────────────────────────────────┘

Where Your Webapps Sit Today

Your Stock Research Assistant and BA Assistant are at Level 1 or Level 2. Here's the proof:

YOUR STOCK RESEARCH ASSISTANT (Level 1-2): ┌─────────────┐ ┌─────────────┐ ┌───────────┐ │ Streamlit │───▶│ Pre-fetch │───▶│ Display │ │ UI (input) │ │ + LLM call │ │ response │ └─────────────┘ └─────────────┘ └───────────┘ - YOU fetch the stock data (yfinance) - YOU construct the prompt - YOU send it to DeepSeek - DeepSeek generates text (1 call, no decisions) - No loop. No tool selection. No autonomy. YOUR BA ASSISTANT (Level 1): ┌─────────────┐ ┌─────────────┐ ┌───────────┐ │ Streamlit │───▶│ DeepSeek │───▶│ Display │ │ UI (input) │ │ LLM call │ │ response │ └─────────────┘ └─────────────┘ └───────────┘ - Pure Level 1. Just prompt in, text out. - No retrieval, no tools, no loop. YOUR SCAFFOLDED AGENTS (Level 4): ┌──────────────────────────────────────────┐ │ AGENT LOOP │ │ ┌─────┐ ┌─────┐ ┌─────┐ ┌─────┐ │ │ │Think│─▶│ Act │─▶│Obser│─▶│Think│ │ │ └─────┘ └─────┘ └─────┘ └─────┘ │ │ └────────────────────┘ │ └──────────────────────────────────────────┘ - ba-work-agent and stock-research-agent are Level 4 - LLM decides which tool to call - This is where the market is heading

The Critical Distinction: Workflows vs Agents

This is the #1 thing hiring managers test for. If you confuse these two, you'll fail the interview question "So what's the difference between an agent and a chatbot?"

Anthropic draws a clear line:

WORKFLOW AGENT ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────────────────┐ │ Step 1 │─▶│ Step 2 │─▶│ Step 3 │ │ LLM decides what │ │ (fixed) │ │ (fixed) │ │ (fixed) │ │ to do next │ └─────────┘ └─────────┘ └─────────┘ │ │ │ ┌─────┐ ┌─────┐ │ YOU define the order. │ │Think│─▶│ Act │ │ LLM executes each step. │ └─────┘ └──┬──┘ │ LLM cannot deviate. │ ▲ │ │ │ └───────┘ │ EXAMPLE: │ (loop until │ Your app calls DeepSeek │ goal met) │ with a fixed prompt → └─────────────────────┘ gets text back. No tools, no loop. EXAMPLE: ba-work-agent reads JIRA, decides what to call next, loops until analysis done.
DimensionWorkflowAgent
Who decides the steps?YOU (developer)THE LLM
Execution modelLinear / branchingLoop (Think→Act→Observe)
Can deviate from plan?NoYes — LLM replans after each observation
Cost predictabilityHigh (fixed # of calls)Low (N calls, variable)
Best forPredictable, repeatable tasksOpen-ended, unpredictable tasks
Example from your appsStock Research Assistantstock-research-agent (scaffolded)

Source: Anthropic — Building Effective Agents

When NOT to Use Agents

Here's what most people get wrong: they think everything should be an agent. Anthropic explicitly warns against this:

Anthropic's Rule of Thumb

"When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all."

"Agentic systems often trade latency and cost for better task performance. You should consider when this tradeoff makes sense."

The Microsoft Azure Architecture Center agrees. They define 3 levels of complexity and say: "Use the lowest level of complexity that reliably meets your requirements."

Source: Microsoft Azure — AI Agent Orchestration Patterns

LevelWhen to useYour example
1. Direct model callClassification, summarization, single-stepBA Assistant
2. Single agent + toolsVaried queries needing dynamic tool useba-work-agent
3. Multi-agent orchestrationCross-domain, parallel specialists, security boundaries(not yet built)

The 5 Workflow Patterns (Preview)

Before you reach Level 4 (full agents), Anthropic defines 5 workflow patterns that are simpler and more predictable. These are the building blocks. You should know all 5:

1. PROMPT CHAINING (sequential steps) A ──▶ B ──▶ C ──▶ D "Generate copy" → "Translate" → "Review" 2. ROUTING (classify then dispatch) ┌─▶ Specialist A Input ─┤ └─▶ Specialist B "Is this a refund? Route to refund agent." 3. PARALLELIZATION (many at once) ┌─▶ Agent A ──┐ ├─▶ Agent B ──┤─▶ Aggregate └─▶ Agent C ──┘ "3 reviewers check code for bugs simultaneously" 4. ORCHESTRATOR-WORKERS (dynamic delegation) ┌─────────────┐ │ Orchestrator│───┐ └─────────────┘ ├─▶ Worker A ├─▶ Worker B └─▶ Worker C "Central LLM breaks task into subtasks dynamically" 5. EVALUATOR-OPTIMIZER (feedback loop) ┌─▶ Generator ──▶ Evaluator ──┐ │ │ └─────◀──── feedback ─────────┘ "Write draft → Critique → Revise → Repeat"

Source: Anthropic — Building Effective Agents

We'll build each of these in upcoming lessons. But first, you need to understand Level 4 (agents) deeply — because that's where your scaffolded projects live.

The one-sentence summary

An agent is an LLM that runs in a loop, decides which tools to call, and keeps going until it achieves its goal — everything else (workflows, prompt chaining, routing) is just YOU pre-scripting the steps the LLM follows.

Practice Drill: Map Your Apps to the Spectrum

Open a terminal and go to each of your projects. For each one, answer: What level is this?

  1. cd ~/Documents/Pythonproject/Touseef_Project_Work/ba-work-agent — Read main.py. Find the line where the LLM decides which tool to call. What level is this?
  2. cd ~/Documents/Pythonproject/Touseef_Project_Work/stock-research-agent — Read main.py. How many tools does the agent have? What level is this?
  3. Think about your Stock Research Assistant webapp. Could you make it Level 3 (workflow) by adding 2 sequential LLM calls? What would they be?
⚡ Quick Check
Q1: Your BA Assistant takes a user's question, sends it to DeepSeek, and displays the response. What level is it?
Show answer

Level 1 (Direct Model Call). No tools, no retrieval, no loop. Just prompt in, text out. This is fine! Anthropic says this is the right solution for simple tasks. Don't over-engineer it.

Q2: Your ba-work-agent reads a JIRA CSV, counts priorities, and suggests top 5 — but the LLM decides which tool to call first. What level is it?
Show answer

Level 4 (Agent). The LLM runs in a loop, calls tools (read_jira, count_by_priority, suggest_priorities), observes results, and decides what to do next. You didn't script the order — the LLM chose it.

Q3: A hiring manager asks: "What's the difference between a workflow and an agent?" What do you say?
Show answer

"In a workflow, I as the developer pre-define the steps — call LLM with prompt A, then call with prompt B, then aggregate. The LLM can't deviate. In an agent, the LLM runs in a loop — it decides which tool to call, observes the result, and decides what to do next until it achieves the goal. Workflows are predictable and cheap. Agents are flexible but cost more and can compound errors. You use agents when you can't predict the steps needed in advance."

Where to Go Deeper

The ReAct Loop
← Back to Deep Expertise Track

Want to see these patterns in action?

Explore the live apps built with these agent architectures.

Explore the Lab →

← Back to Deep Expertise Track