An AI agent has retrieved a complete API collection only when it has followed that endpoint’s continuation contract to its terminal condition, preserved the records needed for the task and recorded any unresolved retrieval failures. A short page does not establish completion. Neither does a fluent summary of the records received.

For an internal report, calculate totals from the collected records in application code. Give the model the verified aggregate and supporting evidence for interpretation. If collection stops early, label the result partial and withhold claims about the full population.

Without that label, a manager receiving the report cannot distinguish an exhaustive review from a retrieval that stopped early.

A short page can still have a successor

Consider a hypothetical agency asking its assistant to review every active client channel for unresolved delivery issues. The connector returns a short list. The assistant examines those channels and reports that all client work has been checked.

The first failure can occur before the model reads a message. If the channel-list endpoint supplied a continuation cursor and the connector ignored it, the assistant has never established which channels belong in the review.

Slack explicitly warns that a cursor-paginated response can contain fewer items than the requested limit while more results remain. For that pagination pattern, inspect response_metadata.next_cursor; an empty, null or absent cursor indicates no further results. Use the individual method’s documentation to establish which pagination pattern applies.

Increasing the page limit leaves the stopping rule unchanged. A larger first page can conceal the defect during a small-workspace demonstration and expose it when the customer adds channels.

The connector must preserve the provider’s continuation contract

A generic instruction to “keep fetching” leaves important behaviour unspecified.

Microsoft Graph’s paging guidance requires clients to follow the complete @odata.nextLink URL until the property is no longer returned. It warns against extracting its token and inserting it into a different request. Graph also documents that a page can contain zero results; an empty page alone is insufficient evidence that enumeration has ended.

Let the API adapter own continuation rather than asking the model to reconstruct it. Store each accepted page before advancing the checkpoint, so a failed next request leaves the retained collection identifiable. The run stays incomplete until recovery succeeds under the provider’s rules. Slack cursors expire, so a checkpoint containing an old cursor does not guarantee that tomorrow’s run can resume from it.

Put a request budget around the loop. Detect a repeated continuation value or a failure to make progress, then stop with a visible exception. A broken provider response must not turn an attempt to be exhaustive into an unbounded job.

These controls fit inside the broader AI agent integration contract. Here the acceptance question is narrower: can the system demonstrate which collection it actually traversed?

Keep collection evidence outside the model’s prose

Fetching every page solves only the provider side of the problem. A tool wrapper can still clip the returned text before the model sees it.

Anthropic’s tool-design guidance recommends pagination, filtering and other controls for large responses, with useful instructions when truncation occurs. The design implication for reporting is to separate durable collection from what fits into a model request.

Maintain a collection manifest in the application. The following is a proposed design, not a vendor-standard schema:

Manifest fieldWhat it establishes
Endpoint, tenant and effective identityWhich authorised source was queried
Filters and collection intervalWhich population the report intended to cover
Accepted page references and unique record IDsWhich records were retained, with duplicates distinguishable
Continuation state and termination reasonWhether traversal ended normally or stopped early
Failed requests and excluded recordsWhere coverage remains incomplete
Model-visible selection and truncation statusWhich evidence the language step actually received

The model should not invent this manifest after writing the report. The collector produces it from execution evidence.

A complete collection can support a deterministic count even if the model receives only selected examples. An assertion that every channel was semantically reviewed requires a further accounting: each eligible channel must have a recorded review outcome or a visible exception. Enumerating channel IDs does not prove their contents were examined.

Likewise, structured output can require a collection_status field without making its value true. Populate execution status from the application rather than asking the model to infer it from a shortened transcript.

Bound the word “all” to the collection you can prove

Exhausting a cursor establishes the end of that traversal under the API’s behaviour. It does not by itself establish a frozen snapshot of a changing organisation.

Records can change during collection, and the effective identity can see only its authorised scope. Where the source supplies snapshot semantics, record them. Otherwise state the collection interval and avoid claiming a point-in-time census. Deduplicate by stable record identifier where available; retain versions when changes matter to the question.

For the agency example, a defensible report distinguishes the eligible client-channel list, the channels actually inspected and the unresolved cases found within that inspected set. If channel discovery or message retrieval failed, the report names the gap. It must not quietly translate “no issue found in retrieved material” into “no delivery issue exists”.

This boundary matters most when the report decides who receives attention. An omitted channel can remove a client from the review queue without producing an obvious error in the final prose.

Test the stopping rule before accepting the report

Use controlled fixtures with known record IDs. These are proposed acceptance tests, not an executed provider benchmark.

FixtureRequired result
Short page with a valid successorRetrieval continues; the final collection includes the successor’s records
Empty page carrying a next linkThe adapter follows the endpoint’s continuation contract
Later page fails or the cursor expiresThe run remains partial or restarts under a documented policy; no full-population claim is released
Same record appears across pagesThe aggregate follows an explicit deduplication/version rule
Wrapper clips evidence after successful collectionThe manifest preserves collection coverage and separately identifies limited model visibility
Normal terminal page and complete downstream reviewThe report is allowed to claim completion within its recorded scope

Compare the retained identifiers against the fixture’s expected set, then inspect the published report. Correct pagination inside a connector is insufficient if the language layer still describes a partial review as complete.

Model Operator is a Dubai-based AI product studio founded by Alexander Chrisostomou. Through AI Initiative Consulting, Alexander works with teams to map consequential workflows and define the evidence their implementation must produce. Bring one internal report and its source API to a Model Operator build conversation; agree what its totals are allowed to claim before expanding its distribution.