> ## Documentation Index
> Fetch the complete documentation index at: https://docs.praxis-ai.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Agents
> Give a Digital Twin its own workspace so it can carry out long-running work on its own — with scheduled tasks, live sessions, and an approval queue you control.
Most of what a Digital Twin does happens inside a conversation: you ask, it answers, the exchange
ends. An **agent** is the same twin given somewhere to work — a private workspace that persists
between conversations, so it can carry out longer jobs, pick up where it left off, and run tasks on
a schedule without anyone sitting in front of it.
This is a separately provisioned capability. A Digital Twin does not have a workspace until an
administrator enables it, and the **Agents** tab only appears once it has been.
## Turning it on
Agent capability is enabled per Digital Twin, in two steps.
In the instance editor, open **Configuration and Integrations → Configuration → Autonomous
Agent** and turn the switch on.
This switch is restricted to Praxis super-admins. If you need it enabled for your Digital
Twin, contact [humans@praxis-ai.com](mailto:humans@praxis-ai.com).
Once eligible, an **Agents** tab appears in the instance editor. This is the control centre —
workspace status, scheduled tasks, live sessions, and anything waiting on your approval.
Activating provisions the workspace. It takes a few minutes the first time, and the status
moves through *Requested* → *Provisioning* → *Ready*.
## The Agents tab at a glance
Across the top of the tab, five tiles summarise the current state:
| Tile | What it tells you |
| --------------------------- | -------------------------------------------------------------------------- |
| **Workspace** | Whether the agent has somewhere to work, and whether it is ready right now |
| **Important Shield events** | Security events worth an administrator's attention |
| **Heartbeat tasks** | How many scheduled tasks exist for this twin |
| **Live sessions** | Work happening right now |
| **Pending approvals** | Actions the agent has paused on, waiting for a human decision |
If a tile reads zero and you expected otherwise, check the workspace status first — scheduled
tasks and sessions cannot run while the workspace is unavailable.
## Workspace status
The workspace is the agent's own machine: its own files, its own home directory, its own failure
domain. It reports one of the following.
| Status | What it means | What to do |
| ---------------- | ---------------------------------------------------------- | ------------------------------------------------------------------ |
| **Not enabled** | The twin is not yet agent-eligible | Enable **Autonomous Agent** (above) |
| **Requested** | Activation received, work not yet started | Wait — this is brief |
| **Provisioning** | The workspace is being built | Wait a few minutes |
| **Ready** | The agent has a working environment and can be used | Nothing |
| **Failed** | Activation did not complete | Read the message shown, then retry. If it repeats, contact support |
| **Stopped** | The workspace was torn down, by you or by an administrator | Re-activate to get a fresh one |
| **Sleeping** | Idle long enough to be put to sleep; files are preserved | It wakes automatically on next use |
**Sleeping is not data loss.** The compute is released to avoid paying for an idle machine, but
the agent's home directory is kept and re-attached when it wakes. **Stopped** is different — a
new workspace starts clean.
## Heartbeat tasks
Heartbeat tasks are the agent's recurring work: things it should do on a schedule rather than
because someone asked. A daily digest, a weekly check over a data source, a periodic tidy-up.
From the Agents tab you can:
* **See every task** defined for this Digital Twin, and whether each is currently enabled
* **Enable or disable** a task without deleting it — useful for pausing a noisy job
* **Create a task** from the available task types
A disabled task stops running immediately, but anything already in flight finishes. Disabling is
not a way to cancel work that has already begun — use **Live sessions** for that.
## Live sessions
Work happening right now. Each entry shows what the agent is doing and lets you **cancel** it if it
is stuck, wrong, or simply no longer needed.
Cancelling stops that piece of work only. It does not disable the task that started it, so a
scheduled task will start again at its next scheduled time — disable the task itself if you want it
to stop recurring.
## Pending approvals
Some actions pause and wait for a human rather than proceeding on their own. They collect here, and
nothing happens until someone decides.
Approve an action and the agent continues from where it paused. Decline it and the agent moves on
without taking that step. Either way the decision is recorded, so there is an audit trail of who
allowed what.
A growing approvals queue usually means the agent is doing exactly what it should — pausing at
the boundary you set. If the same approval keeps reappearing, it is worth asking whether the task
should be scoped differently.
## Important Shield events
Security-relevant events raised by [Praxis Shield](/mdx/admin-guide/praxis-shield) for this Digital
Twin's agent activity — surfaced here so you do not have to go looking for them. Only events worth
an administrator's attention are shown; routine activity is not.
## Troubleshooting
First activation takes a few minutes. If it stays there much longer, the workspace has usually
failed to attach to its runtime. Deprovision it and activate again — a fresh attempt is the
normal remedy. If it happens twice, contact [humans@praxis-ai.com](mailto:humans@praxis-ai.com)
rather than retrying repeatedly.
The workspace exists but the connection to it is not completing. Close the desktop window and
reopen it — the access it uses is short-lived and is reissued each time. If it still fails, the
runtime behind a *Ready* workspace may have gone away; deprovision and re-activate.
Idle workspaces are reclaimed automatically, but not instantly. If one is still listed well
after it went quiet, you can deprovision it directly rather than waiting.
Check three things in order: the workspace status (tasks cannot run unless it is usable), that
the task itself is enabled, and whether the twin still has credits available. A task that
cannot be paid for does not run.
## Related
* [Configuration and Integrations](/mdx/admin-guide/configuration) — where the Autonomous Agent switch lives
* [Praxis Shield](/mdx/admin-guide/praxis-shield) — the security events surfaced on this tab
* [Credit Management](/mdx/admin-guide/credit-management) — how agent work is paid for
* [Entitlements](/mdx/admin-guide/entitlements) — who may administer a Digital Twin