An agent can open the right record and still leave a copy in the wrong place. In its November 16 WorkSpaces announcement, AWS introduced session replay for agent activity. The feature guide labels Replay Session a public preview, disabled by default. Once enabled, it captures the whole screen before each agent action and does not redact sensitive content. That changes the data an implementation team must account for.
AI disclosure: This article was generated with AI assistance from the public sources linked below, with automated fact-checking and editorial inspection. It has not been reviewed by Keith.
A replay is another way to read the record
A recording can be useful when an operator needs to understand why an agent chose the wrong control or missed a field. But the person investigating a desktop problem may not need to read every record visible during that task.
The guide says typed text is also saved in action metadata and can appear in the transcript. Replay Session produces a video, an AI-generated transcript, and an HTML report for delivery to your S3 bucket. Reviewing the video alone therefore misses other places where the same information may appear.
My practical reading is that replay access deserves its own design. The desktop application may have careful role checks, while a broadly shared support folder exposes a recording of what those checks allowed another user to see. An application permission boundary does not automatically describe the audience for an exported recording.
For a proposed deployment, I would separate a routine support view from access to full session content. Support might need a session reference, the failed operation, and a delivery status. A reviewer opening the recording needs a reason and authority appropriate to the content it can contain.
That is an application and operations recommendation, not a claim that WorkSpaces automatically creates those review roles.
Follow the processing path, not just the destination bucket
The documentation describes encrypted AWS-owned processing storage with retention of up to 15 days. Finished artifacts in your bucket follow your bucket’s lifecycle configuration. It also documents Bedrock cross-Region inference: frames and metadata may be processed outside the session’s Region, with routing depending on that Region.
Those details belong in the initial design discussion. “The output lands in our bucket” is not a complete description of where screen content travels or how long every copy exists.
For teams with location requirements, compare the documented processing route with the requirement before enabling the preview. Do not infer inference location from the desktop’s location. Likewise, choose a retention period for delivered evidence based on its operational purpose rather than inheriting the lifetime of an unrelated diagnostics bucket.
Encryption helps protect stored artifacts, but someone must still decide who is allowed to retrieve and view them. A useful implementation names the owner of that decision and makes the access path reviewable.
Keep staged work and real records distinct
The relevant connection to my work is the Primetime Home Health Services case study. It separates a delivered public website from a management implementation whose production data activation, vendor verification, and agency acceptance remain staged. That management foundation includes protected workspace reads, documents, review flows, and audit history.
That distinction matters here. A replay demonstration using fixture records can establish what a proposed recording captures without claiming readiness to record real operational work. The case study does not claim a WorkSpaces deployment, and this article does not propose recording patient information as a convenient test.
Consider a fictional administrative document workflow instead. A synthetic document title, a synthetic contact field, and a synthetic note are enough to trace what reaches the recording. If an unrelated window is visible, give it a harmless marker too. The purpose is to see the capture scope, not to introduce genuine personal data into an experiment.
The same discipline applies to troubleshooting exports. A short clip pasted into a ticket is a new sharing action. Before making that routine, define which content the recipient needs and which approved channel should carry it.
Make evidence delivery part of completion
The guide says processing continues for a few minutes after the session ends. The agent identity must retain the required S3 permissions, and applicable KMS permissions, until delivery finishes; losing them early can prevent delivery.
That creates an operational dependency worth rehearsing. A desktop session ending and its evidence arriving are separate events. A run record should retain the difference, especially if support relies on the recording to investigate a failed task.
For an initial synthetic trial, I would check four things:
- Inspect the video, transcript, and report for the harmless markers placed in the test workflow.
- Try the intended reviewer and a test identity that should not be allowed to open the artifacts.
- End a session and verify the evidence arrives before treating delivery as complete.
- Exercise the chosen retention and support-export process using only those synthetic artifacts.
These are proposed acceptance checks, not tests reported as completed on Keith’s projects. They make the feature’s useful promise concrete: someone with the right access can reconstruct the agent’s activity without turning every session recording into general-purpose support material.
Agent replay can improve investigations. Its adoption also creates a content-access decision. The team should be able to explain who may view a session, where its contents are processed, when delivery is complete, and how the evidence is kept for its intended purpose.
Top comments (0)