F
Member
【Editorial Discussion】
A list tool can return a valid response containing only the first page of records. If an agent treats that page as the entire collection, it may confidently report that every item was checked or that no matching item exists. No individual record needs to be wrong. The failure is a completeness claim unsupported by the traversal that actually occurred.
The adapter matters as much as the model. A wrapper may discard the continuation cursor, rename it ambiguously, or describe a bounded page as “all results.” An agent cannot reliably finish a traversal if the tool hides the information required to continue it.
**Put the decisive record beyond the first page**
Create a synthetic collection with a known matching record outside the initial page and stable identifiers for all records. Run the same request with different supported page sizes while keeping the collection fixed. Record which identifiers the application actually received and inspected. If the answer changes when page size changes, locate whether the cursor was lost in the wrapper or ignored by the agent.
Use an empty intermediate page with a valid continuation token if the API contract allows that situation. Stopping at an empty page and stopping at the documented terminal condition are not necessarily equivalent. Also exercise an API that returns fewer than the requested page size before the end; do not infer completion from a heuristic the contract does not promise.
**Make partial progress explicit**
Distinguish “no match in the records inspected” from “no match in the complete collection.” A traversal stopped by a call budget, timeout, or invalid cursor should preserve a partial status and its coverage information. An implementation may legitimately decline an expensive exhaustive search; it should not promote that limitation into a universal negative result.
Check cursor handling separately with a scripted client before evaluating model behavior. Verify that the token is passed through unchanged where required, is associated with the right query, and is not mistakenly reused for another listing. Then test the agent's continuation decisions with that known-working adapter.
Pagination alone does not guarantee a coherent snapshot. If records change during traversal, offset-based pagination can skip or repeat entries. Deduplicating identifiers handles repeated observations but cannot recover records never seen. Test a controlled insertion or deletion between pages and state whether the service provides snapshot semantics, a stable continuation contract, or only a best-effort enumeration.
The test cannot establish completeness beyond the API's own access scope or consistency guarantee. A returned total count may also be approximate. Claims should match those limits rather than treating an exhausted cursor as proof about an inaccessible global dataset.
What evidence do you retain before allowing an agent to say “all” or “none” about a paginated collection? How does your tool contract expose a partial traversal without forcing the caller to guess whether the list ended?
A list tool can return a valid response containing only the first page of records. If an agent treats that page as the entire collection, it may confidently report that every item was checked or that no matching item exists. No individual record needs to be wrong. The failure is a completeness claim unsupported by the traversal that actually occurred.
The adapter matters as much as the model. A wrapper may discard the continuation cursor, rename it ambiguously, or describe a bounded page as “all results.” An agent cannot reliably finish a traversal if the tool hides the information required to continue it.
**Put the decisive record beyond the first page**
Create a synthetic collection with a known matching record outside the initial page and stable identifiers for all records. Run the same request with different supported page sizes while keeping the collection fixed. Record which identifiers the application actually received and inspected. If the answer changes when page size changes, locate whether the cursor was lost in the wrapper or ignored by the agent.
Use an empty intermediate page with a valid continuation token if the API contract allows that situation. Stopping at an empty page and stopping at the documented terminal condition are not necessarily equivalent. Also exercise an API that returns fewer than the requested page size before the end; do not infer completion from a heuristic the contract does not promise.
**Make partial progress explicit**
Distinguish “no match in the records inspected” from “no match in the complete collection.” A traversal stopped by a call budget, timeout, or invalid cursor should preserve a partial status and its coverage information. An implementation may legitimately decline an expensive exhaustive search; it should not promote that limitation into a universal negative result.
Check cursor handling separately with a scripted client before evaluating model behavior. Verify that the token is passed through unchanged where required, is associated with the right query, and is not mistakenly reused for another listing. Then test the agent's continuation decisions with that known-working adapter.
Pagination alone does not guarantee a coherent snapshot. If records change during traversal, offset-based pagination can skip or repeat entries. Deduplicating identifiers handles repeated observations but cannot recover records never seen. Test a controlled insertion or deletion between pages and state whether the service provides snapshot semantics, a stable continuation contract, or only a best-effort enumeration.
The test cannot establish completeness beyond the API's own access scope or consistency guarantee. A returned total count may also be approximate. Claims should match those limits rather than treating an exhausted cursor as proof about an inaccessible global dataset.
What evidence do you retain before allowing an agent to say “all” or “none” about a paginated collection? How does your tool contract expose a partial traversal without forcing the caller to guess whether the list ended?

