Domain 515% of the exam
Domain 5: Context management and reliability
A bigger context window does not fix lost detail. What Claude Architect Domain 5 (15% of the exam) tests, from caching to recovering a long run.
Questions in this domain keep offering a larger context window as an option, and it does not fix lost detail. Knowing why gets you a long way through Context Management & Reliability. The domain is 15% of the exam and about 9 questions, the lightest domain on the exam, and still enough to decide a result near the pass mark.
The trap is easy to fall into because it matches how we think about memory. If something got forgotten, there must not have been enough space. A model loses detail in three ways. It gets crowded out, buried in the middle of a long input, or compressed away by a summary that kept the gist and dropped the numbers. A bigger context window fixes none of those.
Every module in the domain is a consequence of that one pressure.
| Module | What it's about |
|---|---|
| 5.1a · What fills a context window | What crowds it out, and what to do before anything is compressed. |
| 5.1b · Caching, state and scratchpads | How caching reads a prompt, and where knowledge lives when the conversation can't hold it. |
| 5.2 · Escalation | When a person takes over, and what they need to receive. |
| 5.3 · Error propagation | What a subagent reports upward when it fails. |
| 5.4 · Long exploration | Why an agent grows vaguer over a long run, which a token limit does not explain. |
| 5.5 · Calibration | The checks a team owes itself before it lets a system run with less review. |
| 5.6 · Provenance | Keeping every claim traceable, and handling sources that disagree. |
What takes up the context window
The obvious guess for what fills an agent's context is the conversation. Tool results are what fill it. They are long, verbose, mostly irrelevant, and the first part of the context window to go out of date. The two halves of 5.1 are about dealing with that before it becomes a problem. Trim what a tool returns before it lands. Keep the specifics that must survive (names, amounts, dates, exact wording) somewhere a summary can't touch. And write down anything the agent mustn't lose.
Caching belongs to the same module because it rewards the order you put things in. Put static material first and the parts that change after it. Get that backwards and nothing is ever reused, however much you send.
Failing out loud
Two modules are about handing over badly. Escalation is framed through a support agent talking to a real customer, and the exam is designed to catch the instinct to read how upset someone sounds as how hard their problem is. The module sorts out which situations need a person and which don't. The second failure is the handover itself. A colleague who picks up a conversation cold, with no account of what was tried, starts from zero.
Error propagation (5.3) leans on Domain 2's account of tool errors, and the two read best together. The theme is the same one level up. A coordinator can only work around a gap it knows about, so a subagent that couldn't finish has to say so, and say what it did manage. A report that admits a hole is more useful than one that hides it, even though it looks worse.
Long runs, and numbers you can trust
In long exploration (5.4), an agent early in a session talks about the specific things it found. Hours later it talks in generalities, describing what code like this usually looks like instead of what it saw. That isn't a token limit, so a larger context window doesn't fix it. Writing findings out of the conversation and onto disk does, because they can't fade there.
Calibration (5.5) rewards production experience. An overall accuracy figure can look excellent while one kind of input is close to useless, especially when the common case dominates the volume. A model's own confidence score is worth something only after it has been checked against answers you already know. The module lays out what has to happen, and in what order, before anyone reduces human review, and the exam offers the last step as the first.
The domain ends with provenance (5.6), where research systems lose attribution. Sources get gathered correctly, and attribution gets lost later, at the step that turns findings into prose. And when two sources give different figures, the first job is to check whether they were measuring the same thing at all. A figure for the year and one for a single quarter are not a contradiction. Picking one, or averaging them, throws information away.
What the modules cover
The seven modules cover what to keep outside the context window and in what form, how caching rewards the order of a prompt, the shape of a report a coordinator can recover from, the fields that keep a claim traceable, and the steps before automating, in the order they have to happen. It is a lighter domain by weight, which makes it an easy one to leave until late. Full Access opens the whole domain, and Spaced brings it back mixed in with the heavier ones, on a plan that fits around your evenings, so you are not cramming it the night before.
