Introduction

Agents are a central part of Medbot. An agent reads a patient’s record (or a note, or a form), applies the rules an experienced clinician would follow, and returns structured output fields: a classification, a recommendation, a completed form or a summary. Agents can be run on one patient while you work, or on hundreds in the background, and they can be checked against labelled patients before you rely on them.

How an agent is built and improved

Create it with the four-step wizard: basics, where its patient data comes from, the expert knowledge it follows, and the output fields it returns. See Create an Agent.
Turn its expert knowledge into a workflow. Medbot converts your numbered rules into a graph of steps (gather data, reason, decide yes/no, write the output), checks the graph, and saves it as a version. Only a version that passes every check can go live. See Workflows and Versions.
Bring patient data. Import CSV files or an Excel workbook as a dataset, or connect the agent to a FHIR server or a notes store through an MCP connector. See Patient Data and Connectors.
Validate it against a dataset whose patients carry ground-truth labels, and see how often it is right, field by field and version by version. See Validation.
Run it on patients, notes or forms. See Running Agents.
Collect clinician feedback. Clinicians accept, edit or reject each result. Edited and rejected calls can be added to the validation dataset. See Live Calls and Clinician Review.
Measure and refine. The Results tab shows the agent’s performance, and the Refine flow proposes changes to its rules, workflow and instructions from the cases it got wrong, and saves them as a new version.

The Agents page

Open Agents from the sidebar to go to medbot.me/agents . Each agent is shown as a card. The header has Learn about Agents, the notifications bell, View background runs and + New agent.

Use the chips to choose which agents to show:

Chip	Shows
<your organisation>	Agents owned by your organisation (shown when you are working in an organisation).
All	Every agent you can see.
Mine	Your own agents.
Default	Agents managed by Medbot or your administrator. You can run and copy them, but not edit or delete them.
Public	Agents other Medbot users have listed publicly.
Shared with you	Agents other users have shared with you.

Search with Search agents, and sort by Last run (the default), Last updated, Date created or Name.

Agent cards

Each card shows the agent’s name and description, badges for its scope (for example Mine, Default, Public, Shared by <name> or your organisation), its model and its workflow status, and three tiles:

Tile	What it shows
Validation	The share of labelled patients in the agent’s validation dataset it got right, with the number of patients. Shows No dataset or Not run yet until a validation run has finished.
Agreement	The share of live calls reviewed by clinicians in the last 30 days that were accepted as drafted.
Live calls	How many times the agent ran on live data in the last 30 days, and how many of those were flagged (failed, marked as a false positive or negative, or rejected by a clinician).

The footer shows when the agent last ran and its validation dataset, with Open and either Run or Copy to mine. The ⋮ menu has the actions that apply to the agent: Dashboard (its Results tab), Edit, Run, Share, Copy to mine, Claim for <organisation> and Delete agent.

An agent with no workflow shows This agent runs from its prompt alone. Turn its expert knowledge into a workflow. with a Generate workflow link, which opens the workflow editor.

Workflow status

Badge	Meaning
Workflow verified · v<n>	Version n is live, and runs follow it.
Workflow draft · v<n>	The newest version passes every check but has not been made live.
Workflow needs attention	The newest draft fails a check. Fix it in the editor.
Workflow saved	The agent has a plan saved before verified workflows existed.
No workflow yet	The agent runs from its prompt and expert knowledge alone.

Notifications

The bell in the Agents header shows your unread notices. Click one to open it; Mark all as read clears them.

Notice	When you get it	Opens
<agent> · background run finished	Every patient in a background run you started has finished, for example 42 of 44 patients completed, 2 failed.	The agent’s Agent Runs tab.
<agent> · validation finished: 92% (previously 90%)	A validation run has finished and been scored.	The Validation tab.
<agent> · validation dropped 90% → 85%	A validation run scored 3 or more points below the previous one on the same dataset.	The Validation tab.
<agent> · 10 reviewed calls are not in a dataset	Clinicians have edited or rejected 10 or more live calls that are not yet in a dataset (sent to the agent’s owner).	The Live calls tab.
<connector> · extraction finished	A notes extraction run has finished.	The extracted dataset.

Who can see an agent

Kind	Who can use it
Your own agents	Only you, unless you share them or list them publicly.
Organisation agents	Every member of the organisation can run them. Choose the organisation under Who owns it when you create the agent, or use Claim for <organisation> on a personal or public agent to copy it into the organisation.
Public agents	Every Medbot user can find and copy them. Tick List it publicly so other Medbot users can copy it in the wizard.
Default agents	Managed by Medbot or your administrator. You can run and copy them, but not edit or delete them.
Shared agents	Agents shared with you by email or name. Share your own from the card menu or the agent page (Share). An agent that is already public cannot be shared, since everyone can reach it.

Copy to mine makes your own editable copy of any agent you can see.

Tools

Agents can call tools while they run: Medbot’s built-in actions (such as Send Email, with an optional subject and body template), and the tools of any connector the agent reads from. Choose built-in tools under Actions the agent may take in step 2 of the wizard. Administrators can add tools; see Tools.

Email tool

Configuration	Result
Send Email not selected	No email is sent, even if the agent is instructed to send one.
Send Email selected, no subject or body template	The agent writes the email itself.
Send Email selected with a subject and body template	The email follows your template. The body can use placeholders such as `{{age}}` and `{{gender}}`.