The Agent Engineering Lifecycle
Four phases, weighted identically, and the whole certification is arranged around them. Learn this in the first ten minutes and every later question has somewhere to sit.
Build, Test, Deploy, Monitor
Build is the agent itself: the graph, the tools, the prompts, the state. Test is everything in this course: does it work, is it getting better, how do you know. Deploy is getting it running somewhere real. Monitor is watching it once real people are using it.
Each phase is 25% of the exam. Across 40 questions that is exactly 10 questions per phase, and this course is the only prep course named for Test. The pass mark is 28 out of 40, so you need 70% overall, not 70% here.
That arithmetic is worth stating plainly because it sets your study strategy. Ten questions is a lot to leave on the table and it is not enough to pass on. Spend a quarter of your effort here, and spend it on the distinctions rather than on memorising call signatures.
Why it is a loop
Drawn as a pipeline, this is a waterfall with new vocabulary. Drawn as a loop, it says something true.
The edge that closes it runs from Monitor back into Test. Your agent does something embarrassing in production. You notice it in a trace. You add that run to your dataset. Now it is a permanent offline test, and the next person who refactors that prompt gets a red score instead of a repeat incident.
That harvest edge is where the value compounds. Your dataset is not something you author once at the start. It is a slowly growing museum of every way your agent has ever been wrong, and Module 3 of this course is entirely about the routes exhibits take to get in.
Where observability sits
Observability is not a fifth phase. It is the layer underneath Test and Monitor that produces the data both of them read.
This is why the official course opens with observability and tracing, several lessons before it says the word evaluation, and it is worth resisting the urge to skip ahead. An evaluation score with no trace behind it tells you a number moved. A trace tells you which step took nine seconds, which tool got called twice, and what the model actually saw before it answered wrongly. The number is the alarm; the trace is the diagnosis.
What this course is for
The exam is semi-open-book. During it you may consult docs.langchain.com and smith.langchain.com. You may not use general web search and you may not use an AI.
Think about what that does to question design. Nobody writes a paper where a third of the answers sit one search away in a permitted tab. So the surviving questions are the ones the docs cannot answer quickly: which evaluator type fits this situation, why does this signature differ, what does this number actually license you to conclude. Look-up-able facts get asked as traps rather than as recall.
There is also a LangSmith org provisioned per candidate, and some items are answered by doing something in the interface. So when this course teaches a workflow, learn it as clicks.
Try it yourself
How many exam questions come from this course
Arithmetic worth doing before you decide how many days to spend here. The LCAE is 40 multiple-choice questions, four lifecycle sections weighted equally, pass mark 28 out of 40.
Show answer
Correct answer: D — 10 questions
Four sections at 25% each across 40 questions puts exactly 10 questions in Test. 28 is the tempting wrong answer because it is a real number off the exam spec, but it is the pass mark for the whole paper rather than a section count. The useful consequence: you can lose every question in this section and still pass if you are strong elsewhere, and you cannot pass on this section alone. Treat it as a quarter of your effort, not all of it.
The four phases, in order
This is the spine the whole certification hangs off, and it is the cheapest thing in the syllabus to learn.
Name the four phases of the Agent Engineering Lifecycle in order, and say how the exam weights them.
Reveal answer
Build, Test, Deploy, Monitor. Each is weighted at 25% of the exam, which across 40 questions is 10 questions per phase. This course is the Test phase, and it borrows heavily from Monitor because the two share a surface: everything you observe in production is raw material for the next round of testing.
Why it is drawn as a loop
Which statement best captures why the lifecycle is a cycle rather than a sequence of gates?
Show answer
Correct answer: A — Because Monitor produces the failures that become the next round of test cases
The loop closes because production failures are the highest-value source of new test cases, so Monitor feeds Test. Option 1 describes a waterfall and is the mental model most people arrive with, which is exactly why it is the distractor: phase gates are the thing the loop framing is arguing against. Option 2 is true but incidental, and option 3 is an org chart, not a lifecycle.
Observability is not one of the four
A structural point that explains why this course opens where it does rather than with evaluation.
Observability is not listed as one of the four lifecycle phases. What is its actual relationship to them?
Reveal answer
It is the substrate underneath Test and Monitor rather than a phase alongside them. Tracing is what produces the data that both offline experiments and online monitoring read from, so an application with no tracing has nothing to evaluate and nothing to watch. That is why the official course teaches observability first: evaluation without observation is scoring outputs you cannot inspect, which tells you a number moved but never why.
State the loop without notes
Close the lesson and say the cycle out loud, naming what crosses each boundary.
Something equivalent to build the agent, test it offline against a curated dataset before deploying, deploy it, monitor it online against real traffic, then harvest the interesting and broken production runs back into the dataset and build again. The part that matters is naming the harvest, because that is the edge that makes it a cycle rather than four boxes in a row.