Session workspace
Use the Agent session workspace to follow work, provide input and inspect the result.
A session keeps one piece of Agent work together: the conversation, visible activity, files, questions, feedback and result.
The exact controls can vary with your Project access and the capabilities available to the session.
Conversation and activity
The centre of the session can contain:
your messages and the Agent's replies;
expandable activity while the Agent works;
a run timer and stop control while work is live;
questions or choices that need your input;
files, cards and other outputs;
the final result.
Activity helps you understand the route the Agent took. The resulting business-system state or output remains the strongest proof of what actually happened.
The composer
Use the composer to:
type your next instruction;
attach files;
use speech-to-text dictation;
invoke an available Skill explicitly with /skill;
choose from model options available to you;
answer when the Agent is waiting for input.
Dictation prepares a text message that you can review before sending. It is separate from Call with Agent.
Questions and choices
The Agent can pause with a structured question when it needs information or judgement.
A question can contain selectable options and a free-text answer. Skip means you are choosing not to provide that input and allowing the Agent to continue with what it already knows; that is different from rejecting a required approval.
An Agent question is part of the current task. It is different from a Workflow Request assigned through the Request Inbox.
Files and outputs
Files created or used during the task can appear directly in the conversation and in the session's Files view.
Supported files can open in an inline preview. Open important outputs and inspect them directly.
A file appearing in a session does not automatically make it maintained Project knowledge or a system of record.
See Working with files.
Progress, Context and Memory
For longer work, the right-hand rail can show supporting views around the task.
Progress gives a compact view of the run. It is useful for following work, not proof that every downstream action completed.
Context helps show tools and referenced material around the session. An empty Context view is not proof that nothing contributed.
Memory shows persistent context that may contribute beyond the current conversation. Remembered information is useful context, not automatically the authoritative business record.
Give feedback on a response or the whole session
Use response feedback when your judgement is about one specific completed answer. Responses can show How was this response? with thumbs up, thumbs down and an optional comment.
Use session outcome feedback when your judgement is about the piece of work as a whole. You can return later to edit or remove your own session feedback.
Feedback records human judgement of the result. It does not rewrite technical execution status. A technically successful run can still be a bad business outcome.
Authorised builders can review response and session feedback alongside execution evidence in Monitoring.
Failed turns can be retried
A persisted failed Agent response can expose Retry, letting you regenerate from the failed point instead of reconstructing the conversation manually.
Use Retry for a failed turn. If the same underlying problem keeps recurring, fix the relevant Skill, Connection, context or execution path rather than normalising repeated retries.
Use the workspace as evidence, not the system of record
For important work, verify in this order:
the actual business-system result or required post-condition;
produced files, records and other concrete outputs;
visible Agent activity and the sources it used;
Monitoring or Activity evidence where deeper inspection is needed;
the session rail as supporting context.
The workspace makes Agent work inspectable. It does not replace the systems that own operational truth.
Last updated
Was this helpful?