AnthropicClaude Certified Architect – Professional2026-103.8 Evaluate progressive discovery vs. monolithic context strategy
Choose what every request must load
Completed by the knowledge check
Learning objective
Contrast monolithic context loaded up front with progressive discovery, and design hybrids that preload stable essentials and discover the rest on demand.
Core idea
A small tool menu and concise results reduce irrelevant context. The architectural decision here is broader: which information must already be present before Claude starts, and which information can wait until a task establishes its relevance?
A monolithic strategy supplies the intended tool definitions and reference material up front. Progressive discovery supplies orientation and retrieval capabilities, then loads detail as the investigation develops. A hybrid deliberately combines them. Choose the split for the workload rather than treating either complete preloading or complete discovery as a universal target.
Anthropic’s context-engineering guidance describes just-in-time context as lightweight identifiers, such as paths and links, followed by runtime loading through tools. It also describes hybrids that retrieve some material up front for speed and explore further as needed. Your decision must account for the information Claude can miss, not only the tokens it can avoid.
How it works
Classify information by use and consequence
Separate material into three groups: guidance that governs every request, frequently used operational context, and specialized detail. This classification is an application design, not an API setting.
Put the standing instructions in the Messages API’s system field. Keep frequently needed client tools non-deferred in tools. For rarely needed tools, defer_loading: true permits on-demand definition loading through tool search; see 3.8.2 for the disclosure interfaces. A compact index can identify document families without including their complete contents.
Keep rules that must govern every task available before discovery begins. A restriction hidden in a specialist runbook can be missed precisely when it matters. Always-loaded prose gives Claude guidance; execution controls still enforce the restriction. See 3.8.4 for preserving that boundary when capabilities load dynamically.
Frequency alone is insufficient. An emergency procedure might be rare but essential to a time-critical path. You can preload its trigger conditions and the exact reference needed to obtain it, or preload its full text if there is no time to fetch it. Decide from the consequence of a miss and the allowed acquisition delay.
So material is placed by whether it governs every request, how often it is needed, and, for rare material, the consequence of a miss and the allowed acquisition delay.
Choose the scope of the monolith
Monolithic does not have to mean the organisation’s entire information estate. A role-specific workflow with six tools and two short policies can preload its complete approved working set. The relevant question is whether most tasks need most of that set. Anthropic recommends tool search at 10 or more tools or above 10,000 definition tokens, and standard calling below 10 tools, when every tool is used in every request, or below 100 definition tokens total; these are selection guidelines, not limits (see 3.1.3).
Preloading removes the discovery step for supplied material. It does not establish that the material is complete, fresh, or correctly used. A summary may omit the exception that a decision requires, and a document can be present without receiving sufficient attention. Do not report zero missing-information risk merely because nothing remains to be fetched.
A stable monolithic prefix can also be cached. Prompt caching reuses a matching prefix; it does not remove that prefix from Claude’s context. See 2.5.1 for cache mechanics and 3.8.3 for comparing costs across loading strategies.
Draw a deliberate hybrid split
Use observed task coverage to select the core. Record which operations appear in ordinary requests, which documents supply mandatory decision facts, and which specialist collections are needed only after investigation. A long tail is useful evidence for deferral, but a low call count can also mean that a necessary capability is difficult to find.
Keep orientation small and actionable. An index entry should say what a collection covers and how to obtain its detail. A list of unexplained document IDs saves tokens while leaving Claude to guess what they mean. Anthropic’s just-in-time guidance identifies paths, naming conventions, and other metadata as signals for navigating context.
Choose stable essentials separately from volatile facts. An invariant can remain in system; a daily account status should come from an authorized lookup. Copying live operational state into a reusable prefix makes the preload convenient but creates another version to maintain. Conversely, discovering a stable procedure on every request adds work without necessarily improving freshness.
Assign an owner to the split. The platform team owns request assembly and tool deferral; domain owners own the completeness and meaning of the indexes. Revisit the split when a workflow becomes common, a catalog changes scope, or acquisition delay starts consuming the response budget.
Key distinctions and trade-offs
The strategies divide as follows:
| Strategy | Strong fit | Accepted cost |
|---|---|---|
| Monolithic | Small approved set, broad use on most tasks, little discovery time | Irrelevant material remains present; updates affect the shared input |
| Progressive | Large or varied collections, narrow task-specific use | Search and fetch steps, discoverability failures, variable acquisition paths |
| Hybrid | Frequent core plus specialist detail | Two loading layers and an owned decision about their split |
A larger context window changes capacity, not relevance. Anthropic’s guidance emphasizes selecting the smallest high-signal context for the task. A stronger model or a larger window can be a candidate in evaluation, but neither establishes that an oversized working set is the right architecture.
Architecture / decision scenario
A bank’s internal case assistant has 48 approved tools and 90 operating procedures. Its interactive workflow has an eight-second p95 response target. In a representative evaluation, 82% of cases use four lookup tools and one common case procedure; the remainder need specialist investigations. These are scenario measurements, not Anthropic performance claims.
The tempting proposal is to preload every definition and procedure under prompt caching. That avoids discovery delay, but it also carries the whole specialist corpus into ordinary cases. The opposite proposal, discovering even the common procedure and standing restrictions, adds acquisition steps to the dominant path.
The architect chooses a hybrid: standing guidance in system, the four common tools and shared procedure up front, and a compact specialist index. Specialist tool definitions use defer_loading: true; procedure detail is fetched by an application-defined authorized reader. The banking operations owner approves the core’s coverage, while the platform owner measures specialist p95 and missed-context failures.
If specialist cases exceed eight seconds, the team evaluates preloading their trigger-specific essentials or using a separately agreed slower workflow. It does not declare the hybrid successful from the 82% common path alone. The decision preserves the fast ordinary path while making the tail’s additional cost explicit.
Common exam traps
- Moving to a larger context window or model instead of deferring the long tail.
- Preloading summaries and assuming the decisive exception is included.
- Moving standing restrictions into material Claude might never discover.
- Treating cached context as absent context.
- Selecting the core solely from call frequency without checking unmet needs.
- Treating a hybrid as permanent rather than an owned workload decision.
Where the exam guide differs
The exam guide asks you to evaluate “progressive discovery vs. monolithic context.” Anthropic’s context-engineering documentation describes “just in time” context versus retrieval or loading up front; that post and the Agent Skills overview call staged loading “progressive disclosure.” An exam item using the guide’s terms means the same architectural split: decide what loads before work begins and what loads as the task establishes its relevance. Use the product documentation’s terms when locating the implementation.
What to remember
Preload the material whose relevance and acquisition timing justify it. Discover specialist detail from useful identifiers. A hybrid needs a deliberate core, a reliable path to the tail, and owners who measure both coverage and delay.
Knowledge check
A bank’s case assistant must never disclose an account to a third party without verified consent, in every case type. The rule currently sits in a specialist runbook that Claude fetches on demand, and evaluations show cases where Claude answered before fetching it. Most cases need none of the specialist runbooks, and execution controls already enforce consent on the disclosure tools. Where should the rule’s guidance live to cover every case without preloading unneeded specialist material?
system field, loaded before any discovery beginssystem ; a restriction hidden behind on-demand loading can be missed precisely when it matters. A better index entry still depends on Claude choosing to fetch, and a larger window changes capacity rather than relevance while carrying every specialist runbook into ordinary cases.A grid operator’s assistant needs an islanding emergency procedure in about 0.5% of requests. When it applies, the response must begin within two seconds, and fetching the procedure takes longer than that. The procedure fits on a single page. How should the context strategy treat it?
An HR onboarding workflow uses six tools and two short policies, and every request uses all six tools and both policies. A platform engineer proposes deferring the six tool definitions behind tool search to shrink the initial request. What should the architect choose?
A wealth-management assistant keeps its standing guidance in a cached system prefix. A developer proposes copying each client’s account status into that prefix too, because most requests need it. The status changes several times a day. Which choice supplies current status while avoiding a second maintained copy of that changing state?
References
- Anthropic, Effective context engineering for AI agents
- Anthropic, Tool search tool
- Anthropic, Prompt caching
- Anthropic, Agent Skills overview