Agents
An agent is an AI worker with standing instructions. It starts on a schedule, on an event from one of your systems, or when you run it yourself. It works through its instructions against your live data, decides what matters, and notifies your team through your channels.
Where chat answers one question and a report delivers the same structure on a schedule, an agent evaluates the current data on each run and decides whether anything needs to be done.
How an agent works
An agent takes over recurring work too variable for a fixed rule. A run is not a script: the agent reads its instructions, queries your systems live, and decides what to look up next. This lets you ask for things a fixed rule cannot express: "flag anything unusual", "tell me what changed and why it matters".
It acts only within the permissions you grant, every run is bounded and recorded step by step, and it notifies only when its notify condition is met.
Creating an agent
A draft never runs, and automatic triggers stay paused until you start them.
Describe it in chat
New agent on the Agents page opens a conversation with the agent builder, which drafts the whole agent from your description: trigger, applications, instructions, and notification rules.
Review the draft
The draft appears in a panel next to the conversation, and every change you ask for updates it there. Drafts never run and do not occupy a plan slot.
Triggers
An agent runs when one of its triggers fires:
| Trigger | When it runs |
|---|---|
| Manual | You press Run agent, with optional one-off instructions and files |
| Schedule | Daily, weekdays, specific days, or monthly, at fixed times or intervals, in the workspace time zone |
| System event | A matching event arrives from a connected system, such as a new Apaleo reservation |
| Webhook | Any external system calls the agent's signed URL |
Webhook setup is copy-ready on the agent's page: the URL, a signed curl example, and a Postman collection. Webhook runs are capped per agent (240 per hour by default); calls over the cap are refused with a signal telling the sender when to retry.
Instructions, tools, and notifications
The Agent tab shows the agent as four questions: when does it activate, what can it access, what does it do, and who does it notify. Everything is editable in place.
- Instructions are plain language: what to examine, how to judge it, what to produce. Workspace-wide AI instructions apply on top.
- Applications: only the connections you attach, each summarized at a glance as read-only, acts with your approval, or acts on its own.
- Tools: a Python workspace for real computation and file work. It never holds your credentials.
- Skills: which documents the agent can produce: Excel, PDF, Word, PowerPoint.
- Reports as data sources: deterministic, structured numbers to reason over, the same figures your team sees.
- Channels and a notify condition such as "only if something needs attention". Without one, every run notifies. One failing channel never blocks the others.
Permissions and approvals
Safety has two layers: access (the agent never holds your credentials, only a short-lived token limited to the scopes its applications were granted) and permissions (what it may do with that access). Activating a version never changes permissions by itself. Open an application's tile on the Agent tab to set each action to one of three levels:
- Allow: the agent acts on its own.
- Ask first: the run pauses for approval.
- Block: the agent cannot use it.
A run that hits Ask first parks as Waiting for approval. The request lands on the run page, in the Activity feed, and in any channels opted into approval requests. Allow it for this run, allow it permanently, or deny it with a message the agent reads. Unanswered requests expire after 24 hours. Only workspace admins can change permissions.
Runs and history
Every run is recorded step by step. The Runs tab lists them all; opening one shows what started it, every step, approval, and delivery. A running agent can be stopped mid-run.
-
Outcome: at the end of each run the agent states in its own words how it went, with a two-sentence summary and up to three issues, each with a cause such as an expired credential or missing data.
- Completed the run did what the instructions asked.
- Warning the work was done, but something needs a look.
- Failed the run did not deliver, for example because a system it needed was unreachable.
The Runs tab shows the distribution across recent runs. This closing step is billed like any other turn and can be switched off per agent on its Limits tab.
-
Failures are explained in plain language, with what to do next. Actions completed before a stop are not rolled back; the trail shows what changed.
-
The Audit tab lists every call to your systems: what was read, changed, or prevented. The workspace-wide Activity feed shows everything across agents and reports.
Versions
Every builder checkpoint is kept as a numbered, immutable version, and the active one is the behavior the agent runs. Open any version to see the agent exactly as it defines it, differences highlighted, and activate it to roll forward or back. Older versions are kept, not overwritten.
Edit settings directly on the Agent tab, or use Edit with AI: the builder proposes a new version for you to review and activate. Duplicate agent copies the active version into a new draft; deleting an agent removes it with its run history.
Budgets and cost
Agents are bounded at two levels:
- Per run: a tool-call budget (60 calls by default) and a time budget (10 minutes by default) limit how much a single run can do.
- Per month: every agent has its own euro spend cap (10 € by default) on its Limits tab, alongside the workspace-wide limit. An agent that reaches its cap is blocked; other agents are unaffected.
Agent runs use AI credit; the builder's work when creating and editing agents is covered by the plan. See Limits & cost controls for how caps behave when reached.