Guide
How to build a health agent your members actually use.
A health agent owns a job: it starts on a trigger, follows steps, uses tools and leaves a record. Here is what to write down before you build one, the guardrails that matter when the people using it are members and patients, and the numbers that tell you whether it works.
Start from one job, not from a model.
A model that talks well is not an agent. An agent owns a job end to end. Pick the smallest job that hurts today and write it as one sentence with a start and a finish. “When a member has not booked the follow-up five days after a visit, find out why and book it or hand it to a nurse” is a job. “Improve patient engagement with AI” is not.
Onboard a member. Collect the few things the first conversation needs, then open the right journey.
Check in between visits. A scheduled message that asks one question, records the answer and escalates only when the answer is worrying.
Get someone to care. Find the nearest pharmacy, clinic or emergency department, say what to bring, or book the appointment.
Answer plan and medicine questions. Read the rules or the label, answer in plain language, name the source, and hand anything unusual to a person.
Run a reading. A 30-second face scan or a questionnaire, then a short report with the reading and what to do next.
Chase the work your team repeats. Claims follow-up, document chasing, call-backs and reminders, each with the record operations and audit need.
Five parts, written down before you build.
Whatever you build it on, every agent has these five parts. Write all five before a line of code or a sentence of prompt and the build stops being a research project. The parts are the same on Health Loops, where each one is a field in the dashboard rather than a week of work.
Trigger
What starts a run: a member message, an event from your store or your system, an API call, or someone on your team.
Instructions
The steps in order, written so a new person on your team could follow them.
Tools
What it may do: your catalogue, a scan, messages and calls, health reference data, member memory, your own endpoints.
Guardrails
What it must never do, what it must hand to a person, and what it must say instead.
Record
Every trigger, step, tool call and error, kept for operations, clinical reviewers and auditors.
The guardrails that decide whether you can ship it.
Health is where a demo and a production agent differ. These six rules are what a compliance lead, a clinical reviewer and a member ask about, in that order.
It answers, it does not diagnose. Wellness guidance and plain-language explanation, never a diagnosis, a screening result or a prescription, in the instructions and in the copy the member sees.
Every answer names its source. Plan rules, medicine labels and reference ranges come from a named public source the agent can show. An answer without a source is a guess.
A person is one step away. Red flags, self-harm language, pregnancy questions and anything the instructions do not cover go to a human with the context attached.
The member controls the data. Consent at the first message, a way to delete, POPIA and GDPR duties met, and an employer that sees totals and never chats.
No training on member data. The model provider is told, in the contract and in the configuration, that conversations are not used for training.
Every run is on record. Who, when, what was said, which tool ran and where it went, kept so a complaint can be answered weeks later.
Three routes, and what each really costs.
There is no free option. The choice is where the work sits, and who is on call when a run fails at 22:00 on a Saturday.
Run it on a platform. You write the instructions and pick the tools. The platform owns the channels, the record, the models and the uptime. Fastest to live, least to maintain.
Build it on a framework. You own the model routing, channel integrations, storage, retries, observability and the person on call when a run fails. Most control, months before the first member talks to it.
Hand it to an agency. Someone else builds it one of the two ways above. Fast to start, and you still need a home for it afterwards and a plan for the day the agency moves on.
What the first loop looks like.
Start from a job you already know the answer to, such as a WhatsApp check-in after a visit, and use a ready-made agent as the starting point instead of a blank page. Start free and pick one, or read the journeys teams run.
curl -X POST https://api.ollie.health/loop/$ORG_ID/$LOOP_ID \
-H "Authorization: Bearer $OLLIE_LOOPS_TOKEN" \
-H "Content-Type: application/json" \
-d '{"customerId": "cus_123", "reason": "reorder-due"}'curl -fsSL https://raw.githubusercontent.com/Ollie-Health/Health-CLI/main/scripts/install.sh | sh
health init sleep-agent --channel whatsapp
health code --hosted --prompt "a protocol agent that checks in nightly and hands side effects to a person"In the dashboard the same job is a paragraph of instructions, the tools it may call and the rules it must keep, which is why the writing above matters more than the calls below. Docs, the run API and the embed are at docs.ollie.health.
The five numbers that show it works.
An agent is not finished when it answers. It is finished when these five numbers are good and someone looks at them every week.
Did people start? How many of the people who got the first message replied. A loop nobody answers is a copy problem before it is a model problem.
Did they come back? Messages per member per month, not per conversation. One long first conversation and silence after is a bad sign.
Did it reach a person when it should? Handoffs per hundred runs, and red flags caught. Both should be non-zero: an agent that never hands over is not safe in health.
Did it break? Failed runs, tool errors and messages with no answer, per day. Fix these before you write a new loop.
What did it save? Calls deflected, documents collected, bookings made, minutes per case. Pick the one your operations lead already counts.
Pick the journey closest to yours.
Each of these is a written spec you can copy: the trigger, the steps, the tools and the handoff.
Questions
Do I need developers to build a health agent?
On Health Loops, no. You describe the job, pick the agent and the tools, and test it in a playground before it goes live. Writing your own on a framework needs developers plus someone on call when runs fail.
Which channel should a health agent start on?
The one your members already answer. In South Africa that is usually WhatsApp, because it needs no install and no password. The web and the phone come after the first loop works.
Can a health agent give medical advice?
No. Ollie’s agents give wellness guidance and plain-language explanation, never a diagnosis, a screening result or a prescription, and they hand red flags to a person. Face scan readings are wellness readings and blood pressure is always an estimate.
What does it cost to run one?
Health Loops is free to start and then priced on use: runs, messages, calls and scans. A self-built agent also costs engineering time, model spend, channel fees and the person who maintains it.
How long before the first loop is live?
An agent that answers on WhatsApp and hands red flags to a person can be live the same day on a platform. A self-built one is usually weeks to months before it is safe to put in front of members.
Is this a medical device?
No. Ollie is a wellness product. The agents explain, guide and route to care, and record what happened.
Build the first one this week.
Start free and pick a ready-made agent, or book a call and bring the job you want to take off your team.
