Illustrative workspace investigation with progress, findings, citations and a draft artefact. This is invented example content.
Before you begin
You need Vantage access, workspace and relevant dataset access, permission to submit instructions and save content, and an eligible published agent definition. Select a workspace containing a small dataset relevant to a question you want to answer. It must reach the workspace through a collection. Choose inputs you can inspect yourself. Note the dataset and collection identifiers, contents, processing state and known gaps. If time matters, choose exact start/end dates and a time zone that match the available evidence; omit the period for a timeless question. Replace bracketed prompt fields before sending. If the starting workspace or data is missing, use Kato-assisted setup or the direct workspace lesson. Complete the necessary data preparation before treating the workspace as ready to investigate.1
Check the context and inputs
Read the workspace Summary. Confirm your question and scope. Inspect the selected collection, attached datasets, coverage and processing state.Separate available, incomplete and inaccessible input. Missing access or unfinished processing cannot establish that no event occurred. Direct source investigation needs no world-model graph.
2
Open Explorer and orient
Open Agents, start an agent chat and select a permitted agent definition with suitable instructions, tools and model. This is the Explorer investigation route. Do not choose Simple chat for dataset investigation. Send:
Orient me to [workspace name] and [collection or dataset]. My question is [question], within [scope and optional exact period]. Summarise available evidence, coverage, processing gaps and tools. Inspect only at this stage; leave collection, processing and workspace content unchanged.Compare the answer with your dataset inspection. Resolve an unexpected input or scope before continuing.The agent acts with the submitting user’s authority. Sharing a chat does not combine permissions. Owners manage workspace standing instructions; a chat request does not edit the reusable agent definition.
3
Ask the evidence question
Send:
Answer [focused question] using the permitted inputs and scope we identified. Separate source statements from inferences. Cite the exact source version and relevant passage or locator for each material claim. Keep conflicting accounts and missing information explicit.Check scope and support. Your data may support a finding, a qualified answer or a conclusion that the question cannot yet be answered. No fixed result is expected.
4
Inspect and refine a claim
Open one material claim’s citation. Read the cited content and surrounding context. Confirm its source version and distinguish event time from publication time where relevant.Use Follow a finding to its evidence for the checks. If no claim is supported, inspect the reported coverage limitation. Ask Explorer to correct an overstatement, compare accounts or narrow the question. Unsupported explanations remain hypotheses.
5
Save your investigation note
When you have checked the support, send:
Save a workspace artefact named [note title]. Include our question and scope, the findings we checked with exact evidence references, conflicting claims, hypotheses and coverage or processing gaps. Leave it as a draft for inspection.Saving needs Build mode. If a Plan handoff is pending, review and decide it before implementation.Open the saved version and check content, assets, evidence and limitations. Files produced during a turn belong to the artefact only when included in an explicit save.
6
Stop and return
Use the stop control if execution remains active. Committed messages and saved outputs remain; identify unfinished work before leaving.Reopen the investigation and note. Before a later turn, check current access, coverage and the definition version used. A new published definition can affect the next turn without rewriting earlier messages or evidence.