Workflows vs Agents · AI Agents in Generative AI | Generative AI & LLMs

Two Ways to Build

A workflow is a fixed path written in your code: step 1, step 2, step 3, with LLM calls inside. An agent lets the model choose the path at run time. Most production systems are workflows with a few agent-like parts.

Pattern	What it does	ShopEase use
Prompt chaining	output of call 1 is the input of call 2	draft product description → check tone → translate to Hindi
Routing	classify the request, send it to a specialised handler	refund question → refund prompt; product question → search
Parallelisation	run several calls at once and combine	summarise 50 reviews in 5 batches
Orchestrator-workers	one call plans, others do the parts	plan a festive sale catalogue, workers write each section
Evaluator-optimizer	one call writes, another grades and asks for fixes	write ad copy until it passes the brand rules
Agent	model loops with tools until done	support assistant that looks up orders and creates returns

When to Use Which

Use a workflow when the steps are known in advance. It is cheaper, faster and easier to test.
Use an agent when the number and order of steps depend on the request (open-ended support, research, coding).
Start with one well-prompted call. Add a workflow, then an agent, only when evals show you need it.

Example

A routing workflow. The classifier here is a keyword stand-in for a cheap model call (for example Claude Haiku). The path is fixed in code, so it is easy to test.

def classify(message):  # stand-in for a small, cheap model call
    m = message.lower()
    if any(w in m for w in ("refund", "money back", "return")):
        return "refunds"
    if any(w in m for w in ("where", "track", "delivery")):
        return "tracking"
    return "general"

HANDLERS = {
    "refunds": lambda m: "[refund prompt + refund policy docs]",
    "tracking": lambda m: "[order lookup tool + tracking prompt]",
    "general": lambda m: "[general FAQ prompt]",
}

for msg in ["I want my money back for order SE1042",
            "Where is my parcel?",
            "Do you sell Prestige pressure cookers?"]:
    route = classify(msg)
    print(f"{route:9} <- {msg!r}")
    print("          uses", HANDLERS[route](msg))

Output

refunds   <- 'I want my money back for order SE1042'
          uses [refund prompt + refund policy docs]
tracking  <- 'Where is my parcel?'
          uses [order lookup tool + tracking prompt]
general   <- 'Do you sell Prestige pressure cookers?'
          uses [general FAQ prompt]

Tip

Tip: Anthropic's own guidance is "find the simplest solution possible". Many "agents" in job descriptions are really routing + RAG workflows. Know both and explain the trade-off in interviews.