LLM vs workflow vs agent
All three appear in modern systems, but they are different levels of capability. An LLM (or MMM) is the stateless, passive reasoning core — it responds when prompted but never initiates. An automated workflow executes a fixed if-this-then-that script — deterministic and rigid. An AI agent is a complete system that uses an LLM as its brain to autonomously pursue a goal — it is goal-oriented, stateful, and adaptive.
| Characteristic | LLM | Automated workflow | AI agent |
|---|---|---|---|
| Core function | Generate text from a prompt | Run a fixed task sequence | Autonomously achieve a goal |
| Decision-making | Pattern matching; no decisions | Hard-coded logic; no reasoning | Reasons, plans, decides dynamically |
| State & memory | Stateless | Generally stateless | Stateful; short- & long-term memory |
| Adaptability | Only via retuning/prompting | Rigid; needs re-programming | Learns from experience |
| Failure handling | Silent hallucination | Brittle; fails hard | Resilient; detects & retries |
Key concept · Four characteristics of a true agent
Autonomy (self-governs toward a goal), reactivity/sensing (perceives and responds to its environment), proactivity/goal-orientation (takes initiative, plans over time), and social ability (optional; communicates with other agents/humans via shared protocols). An MMM has rich perception but is not an agent — it stays passive until the surrounding architecture makes it act.
Components as functional building blocks
The same anatomy from Chapter 1 — Goals, Sense, Reason, Plan, Act, Memory, Coordinate — is now read as architecture: goals define the objective function, sense/memory feed context, reason/plan form the cognitive core, act executes via tools, and coordinate handles multi-agent collaboration. Each maps to specific technical challenges.
| Technical consideration | Component(s) affected |
|---|---|
| Data processing & integration | Sense, Memory |
| Knowledge representation | Reason, Memory |
| LLM integration & orchestration | Reason, Plan, Coordinate |
| Reliable tool use | Act |
| State management & memory | Memory |
| Security & governance | Reason (prompt injection), Act (sandboxing), Memory (privacy), Coordinate (authN/authZ) |
How agents talk — and the stack beneath
When multiple agents share an environment, two interaction models govern how they exchange information:
- Direct communication
- Agents explicitly send messages to one another using a common language/protocol (e.g. a structured
reorder_partmessage).
- Indirect (stigmergy)
- Agents interact by observing and modifying a shared environment — one updates a database record, another senses the change and reacts.
Key concept · The three-layer emerging stack
- Layer 1 — Function calling: an agent's LLM triggers a local tool within one runtime. First step beyond text generation; the developer hosts and secures the tool.
- Layer 2 — Tool protocols (MCP): a standard way to discover and use external tools as interoperable services, decoupling hosting from the agent — "tools as portable services, like REST APIs."
- Layer 3 — Agent-to-agent (A2A): collaboration between independent agents across frameworks or enterprises — a universal standard for delegating structured tasks; "SMTP for AI agents."
Many systems are hybrid: function calling for internal actions, higher-level protocols for external collaboration.
Five architectural features, illustrated
Well-designed agentic architecture yields five features — shown here through a supply-chain system:
Modularity
Add a new InvoiceFraudAgent without rebuilding the InventoryAgent; updates touch only that agent.
Scalability
Peak season spins up multiple LogisticsAgent instances behind a load balancer.
Adaptability
A comms agent learns that "production delay" emails precede shortages and auto-escalates them.
Multimodal interaction
A dock agent reads a PDF manifest and inspects pallets by camera before updating inventory.
Collaboration
A disruption agent flags a port closure; the inventory agent messages logistics for alternate routes.
These features — plus the mapping of technical challenges to components — set up Part 2's design patterns, which combine these building blocks to solve recurring problems.
Chapter 4 quiz
Fifteen questions on the autonomy hierarchy, agent anatomy, interaction models and the stack, and architectural features. Answer first, then expand Show answer.
Part A · Multiple choice
What most distinguishes an AI agent from an automated workflow?
- A The agent runs faster
- B The agent reasons, plans dynamically, and adapts, rather than following a fixed script
- C The workflow uses an LLM
- D Nothing — they are the same
Show answer
B. Agents are goal-oriented, stateful, and adaptive; workflows are deterministic and rigid.
Why is an MMM (multi-modal model) not itself an agent?
- A It can't process images
- B It is stateless and passive — it responds when prompted but doesn't initiate actions or pursue goals
- C It has no parameters
- D It only works offline
Show answer
B. The surrounding agent architecture turns passive model potential into active, goal-directed behavior.
Which is NOT one of the four characteristics of a true agent?
- A Autonomy
- B Reactivity
- C Determinism
- D Proactivity
Show answer
C — Determinism. The four are autonomy, reactivity, proactivity, and (optional) social ability.
Indirect communication (stigmergy) between agents works by:
- A Sending explicit direct messages
- B Observing and modifying a shared environment such as a database
- C Fine-tuning each other
- D Sharing weights
Show answer
B. One agent changes a shared record; another senses the change and reacts.
In the three-layer stack, which layer discovers and uses external tools as interoperable services?
- A Layer 1: Function calling
- B Layer 2: Tool protocols (MCP)
- C Layer 3: A2A
- D None
Show answer
B — Layer 2 (MCP). It decouples tool hosting from the agent; Layer 1 is local function calling, Layer 3 is agent-to-agent.
A2A is described metaphorically as:
- A "REST for tools"
- B "SMTP for AI agents"
- C "A load balancer"
- D "A vector database"
Show answer
B — "SMTP for AI agents." A universal standard for delegating structured tasks between independent agents.
Which agent components does the security/governance challenge touch most?
- A Only Goals
- B Reason (prompt injection), Act (sandboxing), Memory (privacy), Coordinate (authN/authZ)
- C Only Sense
- D None
Show answer
B. Security manifests across multiple components, not one.
Part B · True or false
An automated workflow can reason and adapt its behavior to unforeseen circumstances.
Show answer
False. It follows a fixed if-this-then-that path; it lacks dynamic reasoning and adaptation.
A true agent maintains short-term and long-term memory to learn and adapt.
Show answer
True. Statefulness is a defining difference from stateless LLMs and workflows.
Function calling requires the developer to host and secure the tool in the same environment.
Show answer
True. That coupling is its main limitation, which tool protocols like MCP address.
Many real systems use a hybrid: function calling internally, higher-level protocols for external collaboration.
Show answer
True. Hybrid approaches are common in sophisticated agentic systems.
Part C · Short answer
Name the three levels of the "hierarchy of autonomy."
Show answer
LLM, automated workflow, and AI agent — increasing capability and autonomy.
Name the three layers of the emerging agentic stack.
Show answer
Function calling, tool protocols (MCP), and agent-to-agent protocols (A2A).
Name the two agent interaction models.
Show answer
Direct communication (explicit messages) and indirect communication / stigmergy (via a shared environment).
Name the five architectural features of a well-designed agentic system.
Show answer
Modularity, scalability, adaptability, multimodal interaction, and collaboration.