At our recent AI webinar “What’s Next for AI at Canvas”, we demoed three agents live:
All three Canvas agents run on the same platform, coordinate with one another, and keep a human in the loop by design. (For the bigger picture behind the demos, see the two shifts reshaping the EMR.)
Studio: building without code
Studio lets a clinical or operational expert build and customize Canvas by describing what they want in plain language, with no engineering resources required. In the demo, the team built an agent that automatically creates front-desk tasks when a patient without insurance or a point of contact books an appointment, or is a no-show entirely.
Here’s what happens: you describe the requirement in a conversation. Studio asks clarifying questions, then proposes a plan for a human to review before anything is actually built. It builds in your non-production environment and returns a testing checklist, runs the checks it can, and hands the rest back over to you. Once validated, you greenlight it to production.
Studio’s output is reliable because it knows Canvas’s native functionality, reads the developer documentation, and references existing open-source plugins, so it builds patterns that work in your system. Administrators control how builds release to production, whether straight through, after testing, or gated by an engineering review process.
Adoption is the proof: Over 5,000 plugins built and deployed with Studio in roughly five months.
The HIE agent: context before the visit
The HIE agent reads a patient’s external records and reconciles them against what’s already in the Canvas chart, across observations, conditions, medications, allergies, family history, and more. The agent is rolling out in beta, and Canvas is actively looking for customers to co-design it.
For a new patient intake, it can gather a full picture before the first interaction. Its outcome is an up-to-date chart, and it’s billed per completed reconciliation.
The design principle Canvas kept returning to: nothing is written to the chart until a human approves it. Records arrive in clear states: ready to import, needs review, or already in the chart, so the care team can see what exists vs. what’s new. You can run a full reconciliation or focus on one area, like active conditions. For complete records where every required field is present, there’s a direct-add path with no agent review needed. For anything ambiguous, the agent surfaces it to the right person at the right moment rather than guessing.
Reconciliation is genuinely hard, and Canvas is honest about it: the agent handles brand-versus-generic, dosing, and sig differences by flagging them for a care coordinator, not by making unilateral clinical decisions. Getting it to a human at the right time is the feature, not a limitation.
Hyperscribe V2: a scribe that works with the other agents
While scribes are common, what differentiates Hyperscribe V2 is its ability to coordinate with the rest of Canvas. It listens to a visit and proposes structured commands as recommendations, pulling context from the chart, from the note’s own templates, and from data the HIE agent brought in. The clinician stays in control: commands arrive uncommitted, editable, and get added only when the clinician accepts them.
It’s built for real workflow. The scribe docks in a side pane so you can move through Canvas without losing the session, start from your schedule or a chart, and pause for side conversations. Context can be set at the organizational or individual level, so an HPI comes back in the format a given practice wants, and because recommendations are also available via API, a customer can turn off Canvas’s suggestions in the UI and decide how to present the recommendations to its own users instead.
Hyperscribe V2 is in beta, with general availability targeted for the end of 2026. It’s priced the way every Canvas agent is: on the outcome it delivers. For Hyperscribe, that outcome is a finished note, so the unit is a scribed visit, $0.60 each and free through 2026. Canvas CEO Adam Farren published the full rationale and a sourced comparison of 16 other AI scribes if you want to see our pricing methodology.
The connective thread
Notice what runs underneath all three: a human approves before anything hits the chart, each agent draws context from the same shared patient state, and the agents coordinate with one another. When two sources try to write the same observation, Canvas reads the latest state first, so a second diagnosis command gets caught and flagged, whether it came from an agent or a person.
That same coherence shows up in how Canvas prices these agents. Because each one has a single clear outcome, each is billed on that outcome rather than per seat: the scribe on the visit it documents, the HIE agent on the chart reconciliation it completes. You’re paying for the work an agent actually completes.
That’s the Canvas thesis made concrete: the value of any one of these agents is real. The value of running them together, safely, on one platform, is the part that’s hard to replicate.
If you want to see it yourself, the fastest path is to start a Studio project and tinker yourself.
