When should you use a fixed DAG pipeline instead of a dynamic agent loop?

Attempt it

Quick reference

TL;DR

Use a fixed chain when the steps are known up front; use an agent only when the next step must depend on what the last step returned.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Think of it like two ways to cook dinner. The first is a recipe card: chop, fry, simmer, plate. Same steps in the same order every night, so you know exactly how long it takes and what it costs. The second is a chef with no recipe who tastes the dish after each step and decides what to do next. The chef can handle a surprise ingredient the recipe never planned for, but they are slower, pricier, and harder to predict. A fixed chain is the recipe card. An agent is the tasting chef. If you already know every step before you start, use the recipe. Only hire the chef when you genuinely cannot know the next step until you see how the last one turned out.

Key concepts

Agent vs Chain Fixed DAG Pipeline Control Flow

Quick scan · 2–4 min

Everything important, quickly.

Lead with the control flow distinction, name the three guarantees a fixed DAG provides, explain why complexity is not the test, give two tasks where dynamic routing pays off, then close with the per-turn cost penalty and the start with a chain heuristic.

Side-by-side

Property	Fixed DAG chain	Agent loop
Who controls the next step	Developer, at design time	The model, at run time
Latency	Bounded, fixed step count	Grows with trajectory length
Cost per request	Predictable constant	A distribution, hard to budget
Reproducibility	Same input takes same path	Runs can diverge on identical input
Right when	Path is fully known up front	Path branches on intermediate results

Real-world examples

Real products, models, and research that use this idea.

LangChain LCEL chains and LangGraph graphs without loop back edges are the canonical fixed-DAG shape used for predictable production features like classify then route then summarise.
Most retrieval augmented generation features ship as a fixed chain: embed the query, retrieve, then generate, with no run time branching needed.
Cursor and Cline run true agent loops for coding because the next edit genuinely depends on the last test result, which no static graph can encode.
Perplexity's research mode uses an agent loop because each retrieved finding reshapes the next query, the defining case where dynamic control flow pays off.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow do you handle a task that is mostly fixed but has one branch point in the middle?

Sign in or go Pro to see the approach for this follow-up.

QWhat signals during development tell you a chain should be promoted to an agent?

Sign in or go Pro to see the approach for this follow-up.

QHow do cost and latency scale differently between a chain and an agent on the same task?

Sign in or go Pro to see the approach for this follow-up.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.