AnthropicClaude Certified Architect – Foundations2026-105.1 Manage conversation context to preserve critical information across long interactions
Return findings with their decision metadata
Completed by the knowledge check
Learning objective
Require concise upstream findings with claim/source attribution, dates, methodological context, and relevance metadata.
Core idea
A downstream agent cannot reliably restore metadata that an upstream agent discarded. “The program improved outcomes by 18%” lacks the source location, period, population, and measurement method needed to decide what the figure supports. A longer narrative does not fix those omissions unless it actually preserves them.
The exam guide asks for structured upstream data with key facts, citations, relevance scores, dates, source locations, and methodological context. Anthropic’s multi-agent research guidance likewise gives each subagent an objective, an output format, guidance on tools and sources, and clear task boundaries.
See 1.3.4 for separating evidence from source attribution and 1.3.6 for research acceptance criteria. This lesson adds the compact handoff needed when downstream context is limited: each retained claim travels with the metadata that determines its meaning and relevance.
How it works
Put the output contract upstream
In the Agent SDK, use AgentDefinition.prompt to define a specialist’s role and behavior, including the required finding format. Its description explains when to use the specialist; it is not the place to rely on detailed output instructions. Supply the specific question and assignment through the delegated prompt. See 1.3.1 for the broader specialist contract.
Require the specialist to return findings rather than its complete search transcript. Anthropic’s context-engineering article describes subagents conducting extensive exploration while returning condensed summaries, often around 1,000–2,000 tokens. That is an example of the architectural pattern, not an SDK output limit or a mandatory budget for every task.
The SDK’s Subagents page says a non-fork subagent starts without the parent’s conversation; a fork inherits that conversation instead. For the non-fork specialist defined through AgentDefinition here, supply the relevance question and output requirements explicitly. Its task context comes from AgentDefinition.prompt and the Agent tool’s prompt string, alongside the documented project context and tool definitions. Only its final response returns to the parent; do not assume it can infer comparison criteria from unseen turns.
An output instruction is also different from enforcement. Writing “return these fields” in AgentDefinition.prompt defines expected behavior; it does not make missing fields impossible. The receiving application can check that the handoff meets the declared contract before passing it to synthesis. See 4.3.2 for schema-enforced output mechanisms rather than reteaching them here.
Preserve one claim with its attribution
Define one finding per independently supported claim. Keep the source location attached to that finding, with enough evidence to establish why the source supports it. A report-wide bibliography does not show which source backs which sentence.
A compact finding contract can use the following application-defined fields:
| Field | Required meaning |
|---|---|
claim | The supported finding, including material qualifiers |
source_location | Document or URL plus section, page, or other precise location |
evidence | The short passage or extracted observation supporting the claim |
published_at | Publication date, or an explicit indication that it is unknown |
observation_period | The period the underlying measurement describes |
method | How the source obtained the result and its relevant limitations |
relevance | The assigned question or criterion the finding addresses |
relevance_score | A ranking under a declared scale, not a truth probability |
These names are your output contract, not built-in Anthropic citation or research fields. Dates, methods, and relevance metadata come from the exam’s requested handoff design. Anthropic supplies the configurable prompt boundary, while your assignment defines what the specialist must retain.
Keep publication date separate from observation period. A report published in 2026 can analyze transactions from 2024. Labeling its result “a 2026 improvement” would change the claim even though the publication date was copied correctly.
Methodological context should explain what the number measures. A customer survey, a controlled comparison, and a before-and-after operational report support different conclusions. Preserve the measurement unit, population, and material exclusions when they affect the downstream answer.
Make relevance interpretable
A relevance score is useful only with a criterion. For an illustrative zero-to-one scale, define a high score as direct evidence about the assigned comparison, and a low score as background. Require a brief relevance explanation so the downstream agent can distinguish direct evidence from material that merely mentions the topic.
Do not let that score silently stand for source credibility, confidence, or effect size. A source can be highly relevant and methodologically weak. Another can be authoritative but concern a population outside the question. Those distinctions belong in the claim qualifiers and methodological context, not in one unexplained decimal.
Retain decisive contrary evidence even when the majority of findings point elsewhere. The contract should select for the question’s requirements, including exceptions and conflicts, rather than choosing only findings that support a preliminary conclusion. See 1.2.4 for refining incomplete research coverage.
Condense the handoff without erasing its basis
Keep the claim, supporting passage, and metadata concise. Remove repeated narrative, search-by-search commentary, and verbose reasoning chains from what the downstream agent receives. If several findings share a source, a compact source registry can reduce duplication, provided each claim still references the correct location unambiguously.
When required metadata is absent from the source, say so. Do not invent a publication date or method to fill the expected shape. A finding with an unknown observation period may still be useful, but synthesis must know that limitation before comparing it with time-specific evidence.
The coordinator can then assemble a short findings map and labeled evidence sections using 5.1.5. Upstream structure makes that layout possible; downstream rearrangement cannot recover facts that never arrived.
Key distinctions and trade-offs
A concise finding versus a bare conclusion: Concision removes irrelevant exploration. It does not remove the source and qualifiers that make the conclusion usable.
Attribution versus metadata: Attribution identifies where support exists. Dates and methods determine what that support means. A correct URL alone is insufficient for many comparisons.
Relevance versus confidence: Relevance ranks usefulness to a question. It does not certify accuracy or methodological strength.
Prompt requirements versus validated handoffs: The prompt defines the expected result. The consumer checks whether the result actually supplies it before accepting it for synthesis.
Architecture / decision scenario
A research coordinator compares two support automation pilots. One specialist reports an 18% reduction in handling time; another reports 11%. The coordinator initially receives narrative summaries without periods or methods and concludes that the first pilot performed better.
Revise each specialist’s AgentDefinition.prompt to require concise findings with source location, supporting evidence, publication date, observation period, measurement method, and relevance metadata. Include the comparison criterion in the delegated assignment because this non-fork specialist receives its AgentDefinition.prompt and the Agent tool’s prompt string rather than the parent’s turns.
The revised handoffs reveal that the first report excludes multi-concern sessions, while the second includes them, and that their observation periods differ. Both figures directly address handling time, so both can rank highly for relevance without being directly comparable.
The coordinator retains the findings and their qualifications instead of averaging the percentages or declaring a winner. If a source omits its measurement period, the specialist records that gap. The downstream answer can explain the comparability limit and request the missing evidence without spending context on every search the specialists performed.
Common exam traps
- Sending full exploration transcripts to a context-limited synthesis agent.
- Providing a bibliography without claim-level source locations.
- Confusing publication date with the period measured.
- Interpreting an unexplained relevance score as confidence in truth.
- Inventing metadata to satisfy a required output shape.
- Assuming a non-fork specialist sees the parent’s question without an explicit assignment.
What to remember
Define the compact finding contract before research begins. Keep each claim attached to evidence, source location, dates, method, and interpretable relevance, then check the received handoff before using it in downstream synthesis.
Knowledge check
A web search specialist returns long narratives that include its reasoning but omit claim-level source attribution and dates. The synthesis agent, which has a limited context budget, runs out of room and cannot tell which figures come from which source or year. Which change addresses both problems?
A report published in 2026 analyzes support transactions from 2024. A specialist’s finding will be compared with results measured in 2025. How should the finding record the dates?
published_at as 2026 and observation_period as 2024published_at as 2026 and observation_period as 2026published_at as 2024 and observation_period as 2024published_at as 2026, with no observation_period fieldA coordinator delegates to a non-fork specialist defined with AgentDefinition, whose prompt sets the required finding format. The specialist’s findings are well formed but off target, because the comparison the parent discussed earlier never reached it. Which change supplies that comparison?
description fieldrelevance_score field to the required finding formatAgentDefinition.prompt and the Agent tool’s prompt string. The description tells the parent when to use the specialist, the coordinator’s prompt stays with the coordinator, and a score field cannot rank findings against a criterion the specialist never received.A specialist finds a customer survey of 40 respondents that directly measures the handling-time question it was assigned. The survey relies on self-reported times. The contract defines high relevance as direct evidence about the assigned question, separate from methodological strength. How should the finding record this?
methodmethodmethod where synthesis can weigh them. Leaving it out drops relevant evidence, and folding method strength into the score hides it.A source states a result but never says what period it measured. The finding contract requires observation_period. What should the specialist do?
observation_period with the source’s publication yearobservation_period as unknown and keep the findingobservation_period with the year the search was runReferences
- Anthropic, Subagents in the SDK
- Anthropic, How the agent loop works
- Anthropic, How we built our multi-agent research system
- Anthropic, Effective context engineering for AI agents
- Anthropic, Claude Certified Architect – Foundations Exam Guide