Fixed chain or Claude decides?
Some work has steps you can write down before you start. Other work only reveals its next step once you see what the last one found. The trap runs both ways: an open investigation gets a rigid script, or a known checklist gets handed to an agent to improvise. A third wrong answer stuffs everything into one big prompt.
D1 Agentic ArchitectureExam guide 1.6About 15 minutes
Start here: the recipe and the detective
Use a recipe for the first job and a detective for the second. With Claude, the recipe is prompt chaining: a fixed sequence of prompts, where each one handles one piece and the output feeds the next. For a code review, that means one focused pass on each file, then one integration pass that reads all the file notes together and looks for problems between files. The detective is adaptive decomposition: Claude is given the goal and the evidence, and it creates each next subtask from what it has just discovered.
The third shape, one big prompt with everything in it, is not a plan at all. It works for a small job. As the prompt grows, the model's attention spreads thin (the exam guide calls this attention dilution), so later items get less care and the answers can contradict each other.
- Every step is known (review every file, fill every section, the same checks each time): a fixed chain. Each item gets its own pass, then one integration pass looks across them.
- The next step depends on the last finding (an unknown cause, an open investigation): let Claude decide, generating subtasks from what it discovers.
- The job is one small item: one prompt is enough. Splitting pays only once a prompt gets big enough to dilute attention.
Words in the question that tell you which side you're on
Fixed chain
- every file, each section
- “the same steps each time”, “known stages”
- “review each file, then…”
- predictable, must be replayed or audited
Claude decides
- “depends on what it finds”
- open-ended investigation, unknown cause
- “follow the leads”
- a fixed plan finished but missed the real cause
Try it: play out the job
Pick a job, then pick how you split the work. The simulator plays out the steps that run, marks what was missed in pink and what was wasted in grey, and gives a verdict. The diagram above highlights the shape you picked.
The jobs, files and findings are illustrative, made up to show the shape. They are not measurements. The exam guide's point is that known steps suit a fixed chain (each file on its own, then a cross-file integration pass), and that an open investigation should generate subtasks from what is discovered.
Things to try: pull request + one big prompt to watch later files get skimmed, then + fixed chain. Then pull request + Claude decides: it mostly works, at a cost. Then nightly job + fixed chain, and + Claude decides to see why the detective wins there. Last, config file + one big prompt: for one small file, simple is right.
The four traps
| Trap | What the wrong option looks like | Why it loses |
|---|---|---|
| Overbuilt | “Give one agent everything and let it decide the steps”, for a job whose steps are known | It works, but it adds unpredictability and cost for no gain, and a run can skip what a checklist would not. |
| Too rigid | “Hard-code the investigation steps as a decision tree”, or add more steps to a fixed plan | It works for the cases you foresaw. An open problem needs the next step chosen from what was just found. |
| Attention dilution | “Process all the files in one large prompt”, or “use a model with a bigger context window” | Fine for a few files, but quality drops as the prompt grows. Per-file passes plus an integration pass keep each one sharp. |
| Chain without the join | Each file reviewed on its own, with no step that looks across them | Bugs that live between files are never seen. The integration pass is half of the pattern. |
Practice: three exam-style questions
Answer first, then open “Think it through”.
Rule to remember
Exam guide 1.6: “Fixed sequential pipelines (prompt chaining) vs dynamic adaptive decomposition based on intermediate findings.” “Prompt chaining: analyze each file individually, then a cross-file integration pass.” “Adaptive investigation plans generate subtasks from what's discovered.” Exam guide 4.6: “Multi-pass: per-file local analysis + cross-file integration passes (avoid attention dilution, contradictions).”