Every agent has its own computer.
Each agent works in an isolated cloud sandbox: Linux with Python, Node and git, and a /workspace folder for its files. It can run commands, install tools, start servers and test its own work.
Agents can't see each other's sandboxes, your other accounts, or anyone else's work.
It remembers through two files.
BRIEF.mdis yours: what you're building and what matters. The agent reads it before every task.NOTES.mdis the agent's: what it did, what it learned and what's next. It picks up from there after a break.
Edit either from Files, or just tell your agent "add this to the brief".
It sleeps when idle and wakes in a second.
After about 15 minutes with nothing to do, an agent's sandbox goes to sleep. Message it and it wakes in about a second, with everything as it left it: files, installed tools and its conversation.
An asleep agent costs nothing.
Give each agent its own job.
Start another agent from Team › New agent. Each has its own sandbox, files and conversations, so they don't step on each other.
On free credit one agent works at a time. After your first top-up, up to five work at once. Starting one beyond the limit puts your least recently used agent to sleep; nothing is lost.
Tidy up from the agent's ⋯ menu.
- Rename it after the job it does.
- Download its workspace or export the conversation.
- Wipe empties its workspace and conversation so it starts fresh. You type its name to confirm.
- Delete removes the agent entirely.