An assistant recommends an action. The user asks why. A longer explanation appears, but the underlying question remains unanswered: what information did the system use, what did it actually do, and can I correct it? More generated prose is not necessarily more transparency.
For a multi-agent product, I think of a glass box as an inspectable workflow. It should expose evidence and consequential actions without implying that a generated explanation is a faithful record of the model's internal reasoning. This is a design argument, not a report of a user study.
Start with evidence the user can verify
For a research assistant, an evidence view might show the document title, relevant passage, publication date, and retrieval time. Those details answer different questions. Publication date describes the source; retrieval time says when the application accessed it. Neither proves that the source is authoritative or current for this decision.
Attach evidence to the claim it supports. A bibliography under a long answer forces the reader to reconstruct that relationship. If two sources disagree, preserve the disagreement rather than hiding it behind a confident synthesis. If a source cannot be shared with the current user, the explanation must not leak its contents.
Separate actions from explanations
A useful execution history records which component ran, which tool it called, the relevant input and result, and whether the operation succeeded. An explanation can summarize that history, but the underlying structured events remain the source of truth.
In a hypothetical supplier-research workflow, the visible sequence might be: searched the approved catalogue, retrieved three records, excluded one expired record, and drafted a comparison. That is more actionable than a paragraph saying the agents carefully considered all available evidence. It also makes a missing step easier to notice.
Show limits without inventing confidence
A model-generated confidence percentage can look more precise than the evidence justifies. Instead, expose observable conditions: one source is missing, a required field could not be verified, or two records conflict. These are not calibrated probabilities, and the interface should not present them as such.
Where a measured confidence estimate is available, explain what was measured and on which population. Confidence about document classification is not automatically confidence about the final business recommendation. Different parts of the workflow can have different uncertainties.
Make detail progressive, not compulsory
The default view should help the user finish the task. A compact answer can include a few supporting sources and an obvious route to the activity history. Developers may need full diagnostic traces; end users usually need the subset that helps them verify, challenge, or recover.
Do not expose secrets, private prompts, or unnecessary personal data in the name of transparency. Redaction, retention limits, and role-based access belong in the design of the trace system, not as cleanup after logs have accumulated.
Test whether the controls help
Give reviewers realistic tasks containing an incorrect recommendation or missing evidence. Observe whether they can find the issue, explain its consequence, and choose a safe next action. Measure correction and recovery, not merely whether people say the interface feels trustworthy.
Transparency should support appropriate reliance, including the decision not to trust the answer. Pair this design with a usable review workflow and trace-level evaluation. The goal is not a system that persuades people it is right; it is a system they can meaningfully inspect when it is wrong.
An execution record is not hidden reasoning
Expose observable provenance and consequential operations: the source version, tool arguments after validation, state change, and approval outcome. Do not label a generated rationale as the model’s faithful internal reasoning. Evaluate the interface with correct and incorrect cases, measuring false acceptance as well as correction. A more detailed explanation can increase trust without increasing accuracy, so appropriate reliance is the outcome to test.