Skip to main content
AI Agents is where a hospital automates the work nobody enjoys: chasing an unsigned document, answering an out-of-hours enquiry, following up a denied claim, nudging a patient who missed an appointment. You draw the workflow on a canvas, test it in a sandbox, and deploy it — and only then does anything fire.

​
Before you start

Nothing to switch on: AI Agents sits directly under Dashboard and has no care-area gate. On Enterprise you have unlimited published agents, unlimited public chat links, and the webhook trigger — so no plan limit will stop you. Decide first what should set the agent off, because that shapes everything else: a schedule, an event in the hospital (an appointment booked, a claim denied, a lab result ready), an inbound message, or a request from another system.

​
What you’re looking at

The landing page is headed Agent Builder — “Build AI agents and workflows with custom logic and tools.” — with Public UIs, Runs and + Create in the header, and your Agent credits card below: the balance, a Top up button, and a Recent activity ledger where each run appears as Agent run. Your existing agents follow under Your agents as cards, each with a published or draft pill and a Delete link. A hospital with none simply does not get that section; what you see instead is the gallery, headed Create a workflow — “Start from a template, or create your own from scratch.” There are more than thirty templates, and starting from one is much faster than starting from nothing.

​
Inside the builder

Three columns: the node palette, the canvas, and an inspector that says Select a node to configure it. until you pick something. An empty canvas reads “Add nodes from the palette to build your workflow.” The top bar carries the agent’s name, a green ● Live · v{n} pill when a version is deployed, a grey Draft · v{n} pill for what you are editing, a save indicator (Saving… / Unsaved / Saved — there is no save button), your credit balance and the PHI policy switch. On the right: Playbooks, Evals, Preview and Deploy.

​
The palette

Every graph also has a Start node you cannot delete — that is where the trigger lives. Only Agent and Classify call a model, and only those two carry the amber billable chip in the inspector and spend credits. File search, MCP, the logic nodes and the data nodes are deterministic and free.

​
Build and deploy an agent

1

Start from a template

Select a card in the gallery — it installs the whole workflow and opens the canvas. Or select + Create, name it in Create a workflow, and start on an empty board.
2

Choose the trigger

Select the Start node and pick a Trigger. Manual / Preview is the default and is what you want while building. The live options are grouped by Schedule, Webhook / API, Inbound message / call, and then by what happened in the hospital — appointments, documents, billing, memberships, insurance, clinical results, operations, and new patients or contacts.
3

Set the PHI policy

In the top bar, choose 🔒 De-identified · any model or 🏥 PHI allowed · BAA models only. De-identified is the safer default and runs on any model; PHI-allowed restricts you to models covered by a business associate agreement. If the two disagree the inspector warns you and the run is blocked rather than sent.
4

Add your steps

Drag nodes on and connect them. Branching nodes have two labelled outputs — true / false on If / else, loop / done on While, pass / fail on Guardrails — so the path is visible on the canvas rather than buried in a setting.Put a User approval node in front of anything you would not want happening unattended: outbound patient messages, payments, converting a contact.
5

Preview it

Select Preview. The sandbox sheet says Sandbox · remembers this chat; send a test message and watch it run against your real hospital data without the trigger being live. New chat clears the memory. The footer is honest about the cost: “Sandbox — runs the saved agent · debits credits”.
6

Deploy

Select Deploy. You get “Deployed — v{n} is now live”, the ● Live pill appears, and a fresh draft opens at the next version so you can keep editing without touching what is running.
Nothing fires until you deploy, and there is no pause button. A draft has no live trigger, so a schedule or an event never reaches it — this is why a half-finished agent is safe. The moment you deploy, it is running on real patients and real money, and every Agent or Classify step it takes spends credits. There is no undeploy control: to stop a live agent you Delete it. Preview thoroughly before you deploy an agent that sends anything outward.

​
Watch what it does

Select Runs for Agent runs — “Every run is fully traced — status, model, tokens, credits, and per-node steps.” Pick a run to see its Trace, step by step, and its Output. Before there are any it reads “No runs yet. Preview or deploy an agent to see runs here.” If an agent is paused on a User approval node, a Pending approvals banner appears with Approve and Deny on each item. Nothing moves until somebody acts. Evals are repeatable checks against the deployed version — “so you catch regressions before patients do” — including a No raw PHI in output assertion, returning ✓ PASS or ✕ FAIL.

​
Who can do this

Everything an agent touches is still bound by your hospital’s own boundaries — an agent cannot read another hospital’s records, and its runs are logged like any other access. Two behaviours to know: permission enforcement is opt-in per person, and the permissions sheet warns “Saving starts enforcing”; and hiding the row in Role Navigation tidies a sidebar rather than securing anything. A denied action is recorded; an allowed one is not.

​
Check it worked

  • The agent card shows a green published pill and the builder shows ● Live · v{n}.
  • Causing the trigger — booking a test appointment, firing the webhook — puts a new row on Agent runs within a minute.
  • The run’s Trace shows every node it walked and what each cost.
  • Your Agent credits balance has gone down by that amount, and Recent activity shows an Agent run entry.

​
Common issues

The Start node is probably still on Manual / Preview. Change the trigger, then Redeploy — the trigger is only picked up at deploy time.
The agent is set to PHI allowed but the chosen model is not covered by a business associate agreement. The inspector says so in amber: pick a Claude or GPT model, or switch the agent to De-identified in the top bar.
Count the Agent and Classify nodes on the path — those are the only billable ones — and check the trigger. A schedule set to Every 15 minutes runs 96 times a day whether there is work or not. Consider a less frequent schedule with a worklist step.
Look for the Pending approvals banner on Agent runs. A User approval node holds the run until a person selects Approve or Deny — that is the node doing its job.
No server is connected. The node says so: “No MCP servers connected. Add one under Settings → AI → MCP servers, then it appears here.”

​
FAQ

Delete it. There is no undeploy or pause control, so the safe habit is to preview thoroughly and to put a User approval node in front of anything irreversible.
Drafting and editing are free and unlimited. Credits are spent when an Agent or Classify step actually runs a model — including in Preview and in Evals, both of which are real runs.
Only if you allow it. A de-identified agent has identifiers stripped before anything reaches a model, and tool arguments are de-identified before they leave for a third-party server. See AI safety and patient data.
Nothing. Deploying forks a new draft at the next version, so ● Live · v3 keeps running while you work on Draft · v4. Your changes go live only on the next Redeploy.
Yes — publish it as a branded public chat link under Public UIs. Only published, de-identified agents using safe tools are eligible. See Publish an agent.

​
Where to go next

Build an AI agent

The canvas, the nodes and every trigger, in full detail — plus tools and knowledge and publishing.

How agent credits are used

What spends a credit, and what is stripped before anything leaves your hospital — see also AI safety.