Agents

An agent is an AI worker with standing instructions — an automation that reasons, not just executes. It wakes up on a schedule or when something happens in one of your systems, works through its instructions against live data, and notifies your team through your channels when its notify condition is met.

Where chat answers one question and a report delivers the same structure on a schedule, an agent decides. It looks at the current state of your data, judges it against the rules you wrote in plain language, and acts only when acting is warranted.

How an agent works

An agent takes over the work you would otherwise assign to a person: the recurring checking, cross-referencing, and judging that keeps a business running. Watching booking pace, reconciling numbers between systems, inspecting every new reservation for the few that matter — work too variable for a fixed rule, so today it either consumes someone's morning or doesn't happen at all.

When an agent runs, it does not execute a script. It reads its instructions, looks at your live data, and reasons its way to a conclusion — deciding for itself what to look up next based on what it just found. It can chain queries across your connected systems, take the deterministic figures from your reports as trusted input, run real computations in its Python workspace, and produce finished documents — Excel, PDF, Word, PowerPoint — as output. Because it judges rather than matching thresholds, you can ask for things no rule engine can express: "flag anything unusual", "check whether these numbers are plausible", "tell me what changed and why it matters". And each agent carries what it learns from one run to the next, so it gets better at your business over time.

All of that intelligence operates inside a governed frame. An agent reaches only the connections you attach and the capabilities you grant; anything sensitive can be set to require human approval before it happens; every run is bounded by step, time, and spend budgets; and the full trail — every query, every conclusion, every delivery — is recorded step by step. Judgment where rules fall short, without giving up control or auditability.

Where a report is deterministic by design, an agent is intelligent by design, and the two compose: the report produces numbers you can trust, the agent decides what they mean. And because it speaks up only when acting is warranted, a message from an agent means something.

Creating an agent

Agents are built in chat. New agent on the Agents page opens a conversation with the agent builder, which drafts the full configuration from your description: trigger, data access, instructions, and notification rules.

A new agent starts as a draft. Drafts never run, never schedule, and do not occupy a plan slot; they exist so you can review the configuration before anything is live. Activate the draft to make it real. An unfinished draft stays on the Agents page with a link back to the conversation it was built in, so you can resume where you left off, or open its page to activate or delete it directly.

Editing works the same way afterwards: refine the agent in conversation, or adjust individual settings on its page.

Triggers

An agent runs when one of its triggers fires:

• Manual. Run on the agent's page starts a run immediately, useful while testing.
• Schedule. The same schedule builder as reports: daily, weekdays, specific days, or monthly, at fixed times or intervals, timezone-aware.
• Events. Notifications from a connected system (filtered by event type, so a reservation change fires the agent but nothing else does), or a signed webhook URL you can call from any external system.

An agent can be paused at any time. A paused agent keeps its configuration and history but ignores its triggers until resumed.

Instructions, data, and tools

• Instructions are plain language: what to examine, how to judge it, and what to produce. There is no rule syntax to learn.
• Apps. An agent can only reach the connections you attach to it, and within each app only the access you grant. Attaching nothing means the agent works purely from its instructions and reports.
• Reports can be attached as data sources. A report gives the agent deterministic, structured numbers to reason over, the same figures your team sees, instead of re-deriving them ad hoc.
• Workspace. By default each agent has a Python workspace for real computation, and can use document skills to produce Excel, PDF, Word, or PowerPoint output.

Notifications

Choose which channels the agent delivers to and, optionally, a notify condition such as "only if something needs attention". Without a condition, every run notifies. Channels are delivered independently; one failing channel never blocks the others.

Permissions and approvals

Every capability an agent could use is governed by a per-agent policy, visible and editable on the agent's Permissions tab:

• Auto-run: the agent may use the capability without asking.
• Needs approval: the run pauses and an approval request appears in the workspace's approvals inbox (and in channels marked for approval requests). Approve to let the run continue; deny or let the request expire to stop it.
• Out of scope: not covered by the access granted to the agent, so it cannot be used at all.

The Permissions tab also shows the effective permissions: what the agent can actually do once your rules, the app's granted access, and workspace policy are combined.

Runs and history

Every run is recorded step by step: what the agent queried, what it concluded, and what it sent. The agent's History tab and the workspace-wide Activity feed list all runs; opening a run shows the trigger that started it, the full step trace, its duration, and the outcome of each delivery. A running agent can be cancelled mid-run.

A run ends in a clear terminal state: success, failed, cancelled, budget exceeded, or limit reached when a spend limit blocked it before it started.

Budgets and cost

Agents are bounded at two levels:

• Per run: a step budget and a time budget stop a single run from spiraling.
• Per month: every agent has its own euro spend cap (10 € by default) on its Usage tab, alongside the workspace-wide limit. One runaway agent stops itself, not the workspace.

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.