Anthropic CCA-F Questions & Answers

The delegated role is a means of fulfilling the coordinator's task, not a separate authority to change its acceptance criteria. Suppressing citations would produce a result that violates the report's explicit evidence requirement. The coordinator should correct the contradictory brief instead of treating the worker's local instruction as justification for an unsourced deliverable.

Removing citations before synthesis makes precise attribution impossible even if the final prompt later asks for it. Context can be reduced by passing only relevant excerpts and reference identifiers, rather than entire raw documents. Discarding provenance would amplify the very conflict between the specialist's output and the final report contract.

A URL that resembles a legitimate source is not proof that it supports the attached claim. Reconstructing references from memory breaks the chain from retrieved evidence to conclusions. The coordinator should carry actual document IDs and source locations through the handoff and reject any unsupported citations before publication.

Using the current figure may be appropriate for a present-day snapshot, but deleting older provenance prevents the report from explaining change over time. The differing values could be correct for their respective periods. The aggregator should preserve both measurements and their dates before selecting any number for the requested analysis.

A shared company label does not imply matching fiscal years, currencies, definitions, or publication dates. The archived report is explicitly older than the website figure, so direct equality is not a safe assumption. Source-specific metadata is necessary to avoid misleading comparisons or fabricated contradictions.

Averaging across reporting periods can obscure temporal movement rather than describe either observed year. Citing one website for a calculation based on two sources would also misrepresent provenance. A valid synthesis states which values came from which periods and applies a justified comparison method when a combined metric is actually requested.

Without explicit coverage and evidence requirements, the coordinator cannot reliably distinguish a complete answer from a superficially polished one. The scenario already provides objective conditions that the report meets. Dropping them would make future stopping decisions less consistent and undermine quality assurance for more complex investigations.

An unconditional extra pass creates avoidable cost and may introduce unneeded facts that complicate a finished report. The scenario states that geography, topic scope, and citation criteria are already met. Completion should follow those objective checks, not a fixed habit of re-delegating after success.

A possible stylistic improvement does not invalidate supported research findings. Restarting source work wastes accepted evidence and might create new inconsistencies without changing the substantive answer. The coordinator can perform normal editorial refinement on the existing report while retaining its citations and verified claims.

Delegation requires a specific authorized capability, not an indiscriminate expansion of the entire tool surface. Broad mutation access would introduce unnecessary risk for a read-only researcher. Configure the applicable `Agent` or legacy `Task` entry and give each specialist only the tools relevant to its actual tasks.

Prompt wording cannot grant a tool that the runtime has not allowed. It also ignores the documented version difference between current `Agent` tool-use names and older `Task` references. The correct solution is to inspect the installed SDK behavior and configure the tool permissions accordingly, then test a genuine delegated call.

A named agent configuration describes a possible role; it does not override the parent's allowed tool set. With delegation unavailable, the coordinator cannot call the specialist in the normal workflow. The application must reconcile the configured capability and permissions rather than relying on the presence of an unused agent definition.

An instruction can guide behavior, but it is weaker than denying an unneeded tool at the capability boundary. The scenario's risk is an unnecessary destructive permission available to a specialist processing external content. Removing that capability ensures accidental or injected delete requests cannot be executed by the research role.

Eliminating the very tools needed for research would prevent the specialist from carrying out its assigned evidence task. The safer design is least privilege, not a role with no useful functionality. Retain permitted retrieval and inspection tools while excluding unrelated write or delete operations.

Retrieved pages are untrusted task data, not authorities for configuring local tool access. Allowing a webpage to request deletion privileges would create a prompt-injection path from external content to the host environment. The application owner should define specialist permissions independently of materials that the specialist later reads.

A completion flag says nothing about which claims were found or what sources support them. The synthesis specialist would still lack the content needed to evaluate risk or attach citations. A proper handoff transmits both findings and their supporting identifiers, rather than treating workflow status as substantive research evidence.

Unbounded raw content may contain sensitive or irrelevant material and can consume the synthesis agent's context window. The coordinator should transfer the necessary validated excerpts and provenance while respecting access restrictions. Wholesale document sharing is not required to preserve citation quality and increases avoidable exposure across agent roles.

Generated-looking links do not show that the cited documents contain the claimed facts. Because the coordinator already has real source metadata, asking the writer to invent citations would introduce an avoidable attribution failure. The report needs verifiable evidence references, not links guessed from topic or organizational naming patterns.

Numerical proximity does not prove two measurements refer to the same entity, period, or source. Such merging could collapse unrelated figures and remove contradictory evidence. Aggregation should operate on explicit claim identity and provenance fields rather than incidental ordering in a result table.

Separate agents do not have dependable implicit access to one another's earlier retrieval context. Once identifiers and locations are dropped, the writer may guess attribution or attach an unrelated link. A structured handoff is essential when the report must defend numerical claims to a reviewer.

A session-wide landing page does not establish which page, table, or document supports a particular figure. Different analysts may have consulted unrelated sources or reporting dates. Attaching one citation to every value would blur important distinctions and could falsely imply that the generic page verifies unsupported numbers.

Legal evidence is only one part of the requested decision; a regulator-focused worker may not have validated engineering feasibility. Publishing before the technical results arrive would omit a required perspective. The coordinator should remain responsible for combining both analyses into the final product recommendation.

Undifferentiated task briefs can make specialists duplicate work or neglect a needed source family. The workflow benefits from one policy-focused and one implementation-focused investigation. Explicit scopes and output requirements let the coordinator integrate complementary findings rather than reconcile two broad but inconsistent essays.

There is no stated prerequisite linking those two retrieval tasks, so serial execution would increase latency without improving correctness. A regulator's rules can be researched independently of the technical architecture analysis. The coordinator should synchronize before the final recommendation, not before each separate investigation begins.

Excess context can expose unrelated material, while unrestricted mutation privileges exceed the needs of an investigative analysis. The scenario specifies distinct roles and permissions, so identical broad grants erase useful boundaries. Role-specific least privilege allows both analyses without expanding the risk of accidental or injected changes.

A role label is not an adequate task brief when several migrations, schemas, or authorization boundaries could exist. The coordinator has relevant references and constraints that should be passed explicitly. Omitting them increases the risk that the specialists inspect different artifacts or reach conclusions based on incompatible assumptions.

Concurrency should reduce waiting, not eliminate provenance or role separation. If one worker mutates the other's findings, the coordinator cannot tell which source supports a given migration risk. Keep returned evidence and statuses distinct until the integration step, where conflicts can be adjudicated explicitly.

The analysis does not require personal chat messages or financial identifiers, so including them expands sensitive exposure without improving the calculation. Context boundaries are intended to support scoped delegation. A minimal fact packet is sufficient and more defensible than granting a specialist broad access simply because the parent possesses the data.

A summary calculation is read-only analytic work, not authorization to move money. Granting financial action credentials introduces an unnecessary side-effect capability and could permit an erroneous or injected tool call. The coordinator should reserve refund operations for separately guarded support workflows with verified approval.

Privacy protection does not require abandoning the legitimate analysis; the task can use sanitized, purpose-limited facts. Without data, a calculated summary would become speculation rather than an answer to the customer's actual records. The right boundary preserves necessary evidence while excluding unrelated sensitive content.

A tool or agent returning text means it executed, not that the result satisfies the report's requirements. The missing evidence and comparison are substantive quality defects. The coordinator must validate returned findings against the delegated goal and request targeted completion rather than treating message receipt as proof of research success.

More emphatic wording does not tell the specialist which claims need verification or how much coverage the report requires. The existing failure stems from missing task structure, not an insufficiently motivated agent. Concrete evidence fields and completion criteria provide a testable direction that generic encouragement does not.

Parallel duplication could produce several shallow paragraphs without covering the required comparison. Nothing in the scenario indicates that worker count is the limiting factor. The coordinator should first clarify differentiated scope and acceptance criteria, then decide whether multiple independent subtasks are useful.