Your agents are useful because they know what is going on. When a mix finishes and nobody has played it, when an upload stops halfway, or when a release is saved without a date, the agent who owns that work can say so and offer the next step. This page says what they see and what they do not.
What they see
A record of what happened in your workspace, kept for thirty days:
- Pages and things opened, such as a song or a release. Page visits are kept for two days only.
- Uploads started, finished, stopped or failed.
- Things saved: a song, a member, a release, a campaign, a setting.
- Work that started or finished: a mix, a master, a video, an import, a review.
- Failures you saw, as a short code.
- What was played: a song, a mix, a master, a video.
- What the agents themselves did, and what an API key did on your behalf.
Each entry carries who did it, what it was about and when. That is enough for an agent to say “the mix of Alone is ready and you haven’t heard it yet”.
What they never see
The words. Nothing anyone types is recorded there: not messages, not lyrics, not bios, not prompts, not the contents of files. A failure is recorded as a code, never as the message you read. Names you gave your own things, such as a song title, are kept because that is what a sentence needs.
The record belongs to the workspace, like its notifications. Everyone with a seat in the workspace sees the same record, and it says who did what.
When an agent speaks first
An agent only starts a conversation about something that needs you now: a failure in the last few minutes, an upload that has stopped, something left half done, or finished work nobody has played. It opens its own conversation, never writes into one you already have, says each thing once, and never more than one unprompted conversation every half hour. Starting a conversation costs no credits; your reply starts a normal turn.
If you would rather they wait to be asked, set agents to Neutral in your workspace settings or in your own settings. The record is still kept, because it is also how SongPilot audits what agents do.
Ask them
Ask any agent “what did I leave unfinished?” or “what happened this week?” and it answers from the same record. Through the API, the same record is at the activity endpoint, and in Claude the agents can read it with the activity tool.
