News
Healthcare IT News
FeaturedWhy Johns Hopkins is benchmarking AI agents before deployment
Health system leaders say reliable benchmarking, governance and workflow design must come before scaling agentic AI across administrative operations. Johns Hopkins Medicine is taking a deliberately cautious approach to agentic AI, focusing first on proving reliability and governance before expecting measurable financial returns.
Blog
Meet Srinivas Pang, VP of Delivery at actAVA.ai
Srinivas Pang is a Product & Services Executive with 20+ years of experience leading enterprise technology transformation across Healthcare, Financial Services, and Applied AI. More recently, his focus has centered on product and services delivery for health payer organizations, with particular emphasis on Applied AI, Agentic AI platforms, professional services, GTM, customer success, and payer transformation.
Video
ACTAVA | KORA | Agent Connectors (Tools, APIs, MCPs)
Adding a tool is a security decision wearing a convenience costume. Adding a tool to an agent feels like adding a feature. It's closer to granting an account access to a system. The tool list is the blast radius, and it deserves the scrutiny you'd give a service account. It rarely gets it. Ask what it can reach. Read a calendar, or write to a record? Draft an email, or send one? The verb is the entire risk. Ask on whose behalf it acts. Agent credentials, or the user's? One is auditable to a person. The other is a shared account with nobody's name on it. And ask what happens if it's wrong. A wrong search wastes a second. A wrong send cannot be recalled. Reversibility determines where the gate goes. Those questions have a home in the product. Every tool attached to an agent is a door you're opening, and the blueprint lists them together where you can see them. Connectors are the same question one level up: what is this allowed to reach? Access is set per agent, so the answer can be different for each one, which is the whole point. Keys are the same conversation with the volume turned up. And the store is where it becomes a habit, because picking a tool grants permission every time. Ask what it can reach, not what it can do. Features get added. Permissions get granted. Tools are the second thing pretending to be the first.
Video
ACTAVA | Agentic Workforce | Agents That Follow a Plan, Not a Paragraph
Prose describes intent. A plan is something you can check. Most agents start life as a wall of prose. It reads well, and it hides every decision the agent is actually going to make. When written as ordered steps rather than paragraphs, the same instruction changes what everyone downstream can do with it. It becomes reviewable. A colleague reads eight numbered steps and says step four is wrong. Nobody can do that to a page of prose. It becomes addressable. You can mark one step for approval or retry it. There is no middle of a paragraph to mark. And it survives editing, because changing step six leaves the rest of it intact. Rewriting prose quietly changes things nobody noticed. You can see the difference in the product. The work gets defined as a structure rather than a paragraph of hopeful instructions, broken into steps you can read in order, which means you can argue with step three without rewriting the whole thing. Before it does the work, the agent shows you the plan it intends to follow. And a plan you can see as a pipeline is a plan you can change one piece at a time. Which matters, because the ask was one sentence and the work is considerably more than that. So write the paragraph if it helps you think. Then turn it into steps before anyone has to trust it. A plan is reviewable. A paragraph is a hope.
Video
ACTAVA | Agentic AI for Transitional Housing (demo)
Describe the intake workflow. Review what it designs. Keep the counselor in it. If you run intake and counseling for a residential program, you already know where the hours go. Not the work. The writing down of the work. The work happens once, but it gets written down two or three times by the people you hired to do it. That isn't a documentation problem. It's a capacity problem wearing a documentation costume. So this builds one. The whole input is a sentence of plain language: read the referral, check bed availability, assemble a record for the counselor to approve. What comes back isn't an agent that does everything. It's three steps you can read, each with a job you could describe out loud. A referral intake reader. A bed availability checker that fans out to run once per row in the upstream file. And an intake record assembler. Read that third one carefully, because it's the whole point. It assembles a complete intake record for counselor approval, and that is where it stops. It enrolls nobody. The agents prepare the case. The counselor decides it. The created workspace shows the execution flow written out, with one output file per row, and the steps still marked draft rather than released. Designing a workflow is not the same as running it on anyone. Two things decide whether this survives contact with a real caseload: where the rows come from, and where the finished work goes. The source is its own step, fetching rows from a system that already has them so the next step can run once per row. And the connectors are the honest answer on integration. They're the doors that exist. Anything beyond them is a conversation about your systems, not a claim anyone can make on a screen. Then there's the half that decides whether any of this is worth doing: the day being signed off before it counts as billable. A daily run is defined by which agent, which prompt, and what time, and the one here is switched off because defining a schedule and arming it are two different decisions. Alongside it sits the gate that matters: a named group, a policy for how many must agree, and a sign-off on the day before it's submitted for billing. None of this replaces the judgment in the room. It takes the assembling off the people doing the counseling, and leaves the deciding exactly where it was.
Blog
Understanding MCPs and Deep Agents
Model Context Protocol and deep agents are often discussed interchangeably, and they shouldn't be. MCP is an open standard that provides an AI with a governed way to access the systems you already run: the EHR, the payer portal, the population health platform, the CRM. A deep agent is what turns that reach into finished work, taking a goal, planning the steps, executing across live systems, reasoning about what it finds, and stopping for human approval at the decisions that carry consequence. Follow one care manager's Monday through both versions, before and after, and the difference is visible in the first hour: 40 tabs and a printed gap list become one dashboard where every patient has been reviewed, every gap prioritized, and every outreach message drafted for her to check. This piece walks through what each piece actually does, how the ACTAVA Workflow Library maps onto them across HCC review, quality gap closure, and denial management, and why the human checkpoint isn't a concession to compliance but the correct design for clinical AI.
Video
ACTAVA | KORA | Agent Versioning
A prompt you edited is not something you can roll back. Here's a question worth sitting with. If someone changed your agent yesterday and it got worse, what exactly do you roll back to? Editing a prompt feels like editing a document. It behaves like deploying to production, except with no tag, no diff, and no way back. Three different states get collapsed into one when all you have is a prompt box. There's the draft, which is where the work happens. Every save overwrites the last one, so nobody can tell you what it said on Tuesday, or why. There's the version: a snapshot with a note attached saying what changed. That's the moment it stops being something someone edited and starts being something you can point at. And there's promotion, meaning which version actually runs by default. Keeping that separate from the draft is what makes improving an agent and changing what serves people two different acts. You can see all three in the product. The draft contains the system prompt, the plan, and the tasks, but it doesn't record what it said yesterday. Snapshotting it creates a version, and the box asking what changed is the part that matters, because it's the only place the reason survives. Then you have a history: draft and version one side by side, each with who saved it and when, where you can view either, compare them, or restore an old one straight back into the draft. Edit the draft. Ship the version. The question isn't whether you version your agents. It's whether you find out you needed to before or after the rollback you can't perform.
Video
ACTAVA | Agentic Workforce | Agents that State their Limits
Stated limits are a feature, not an apology. There's a strong temptation to describe an agent purely by what it can do. The teams that get trusted do the opposite. A stated limit is checkable. A capability claim usually isn't, which is precisely why the limit carries more weight. It sets the test. "Never submits without approval" is something a reviewer can verify in a single run. "Intelligent" is something nobody can falsify. It survives contact. Overstated capability gets found out on day one of real use. Understated capability is discovered as a pleasant surprise. And it names the owner. Saying what the agent will not decide is the same sentence as saying who does, and that's the one the clinician, or the lawyer, actually needs to hear. So lead with the boundary. It's the part people can check, and checking is how trust gets built.
Video
ACTAVA | KORA | Agent Benchmarking Experiments
Two scored runs side by side, down to the metric, then down to the row that moved. Changing an agent is easy. Knowing what the change actually did is the hard part, and a single overall score cannot tell you. One number against another number is an argument. A named metric and a named row is a finding. Every experiment leaves a scored record behind: the overall number, the agent version it ran against, and a breakdown by metric. A score in isolation means very little. A score next to another run of the same dataset means quite a lot. So put two of them side by side. Comparing two runs produces a scorecard listing every metric with both scores and the delta between them, which means the metrics that didn't move visibly stay put. In this comparison the citation check is identical in both runs, zero delta, while a single metric accounts for the entire difference. Separating the thing that changed from the things that didn't is most of the diagnostic work, and it happens in one glance. Then it narrows again. The per-row comparison scores each case in both runs and shows the difference, and here two rows are identical while one row accounts for the whole gap. That row is a specific case with a specific input. Someone can open it and read it. That's not a number to argue about in a meeting. That's a row you go and read. A score you can't decompose is just a number. A score you can trace to a metric, and then to a row, is a test.
Video
ACTAVA | KORA | Agent Workspace Live Monitor & Approvals
Agents run on their own. This is the queue that tells you which ones need a person. Autonomy is only tolerable if it is observable. Your agents are working right now, so the question isn't whether they're running. It's which of them needs you. The interesting screen is not the one where an agent works. It's the one that tells you which agents are stuck on you. The live monitor lists every run in flight across the organization, and each one declares its state: awaiting you, held for approval, or running. Agents that work while nobody is looking are only safe if there is one place showing you all of them. Nothing is buried in somebody else's browser tab. Open a run, and you get its history, with the duration and the operating-unit cost of every attempt, who started it, and how it ended. Look at what's actually in that list, because it isn't a row of green ticks. Rejected. Rejected again. Awaiting approval. Somebody read what this agent drafted and said no, twice. A gate you cannot audit is a gate you cannot trust, and rejections are the evidence that the review is real. That is what a working gate looks like from the outside. Approver groups are the mechanism underneath. Each one carries a policy, whether that's all must approve, a majority, or the first available reviewer, and a timeout so nothing waits forever on somebody who is on leave. Different decisions deserve different quorums, and none of them should stall indefinitely. An agent that runs on its own is only safe if you can see it and stop it. A queue of what needs you, and a record of what you refused. That's oversight you can actually operate.
Video
ACTAVA | KORA | Agent Dashboard Usage
The model was never the hard bit. Every agent project has a thrilling first week and a difficult second month. The gap between them is entirely made of unglamorous work. None of that second month is about model quality. It's about whether the thing survives contact with an organization. Three things decide it. Who can see it. Access granted per agent to named people. Not everyone with a login, and not a spreadsheet somebody maintains by hand. What it cost, attributed to a team, a project, a person. A pilot with no cost story doesn't get a second quarter. And what changed. Versions, diffs, and a note saying why. Six months in, nobody remembers, and the person who made the change has moved teams. The demo is the easy half. The boring part is the half that's still running next year.
Video
ACTAVA | KORA | Agent Document Themes
Paste the brand guidance you already have, and every PDF, Word doc, and deck an agent exports follows it. There is a specific, unglamorous tax on agent output. Your agent writes a genuinely good report, and then someone spends an hour reformatting it so it looks like it came from your organization rather than a machine. Document themes remove that step. One theme defines how everything your agents export looks, across PDF, Word, PowerPoint, and HTML. The part worth watching is the input. You don't configure a theme field by field. You paste what you already have, as prose. Extract from Brand Docs takes pasted brand guidance or uploaded files, whether that's PDF, DOCX, TXT, CSS, or JSON, and maps it onto a theme. Every organization already has this document somewhere. Retyping it into a form is exactly the kind of work nobody should be doing. Then it reads that prose and builds the theme from it. A primary color, a secondary, an accent, and the hex values come through exactly. Serif headings in sentence case, body text in a named face. The logo lands top left because that's what the text asked for. And the footer line the guidance said must appear on every export becomes the footer. That extraction is an inference, so you check it before you trust it. The theme is presented for review before the card is created, covering colors, logo placement, typography, footer, table, and callout styles. Reviewing it first is the difference between a shortcut and a liability. Alongside it sits a live preview of the export: headings in your color, note, warning, and tip callouts, a styled table, and the footer at the bottom of the page, pulled straight from the guidance you pasted. One theme, four export formats, with per-format overrides for the cases where a slide deck genuinely needs different rules than a report. Consistency across formats is exactly what hand-formatting fails to deliver. Your agents were already writing the content. This is what stops it from looking like they did.
Video
ACTAVA | Agentic Workforce | Agents That Humans Must Approve
Some work an agent does should never reach the world without a person saying yes. Putting a human in the loop sounds like a good idea. It isn't free, and past a certain point it stops buying you safety at all. A gate converts speed into accountability. That's a good trade exactly once per decision that deserves it, and a bad one everywhere else. The real cost isn't the click. It's the wait. Work sits until someone looks, and whoever happens to be on leave becomes a dependency no one wrote down. And there's a worse failure than slowness. Gate everything, and approvals turn into rubber stamps. A reviewer clearing forty items a day is reading none of them. The gate still exists on the org chart, and it has quietly stopped doing anything at all. So the question isn't whether to have approvals. It's where they earn their cost. Put the gate on the irreversible step: sending, submitting, paying, publishing. Let the agent draft as freely as it likes, and stop it at the door that only opens one way. Done properly, that gate is a real control rather than a promise. It's enforced by the system rather than by a prompt asking a model to be careful. It names who decides, and it doesn't wait forever on someone who's away. And it leaves a record, because a gate you cannot audit is a gate you cannot trust. The rejections are the evidence that the review was real. A gate on everything is theatre. A gate on the thing you can't undo is control.
From production workflows to customer-controlled intelligence.
Don't give your agentic future away to a single model provider. Don't mistake consumer tools for real, safe Enterprise Agentic Tools. Enable your citizen developers to create and manage the AI Agents they need to run their part of your business.
Build complex agents. Test their reliability. Learn from every workflow. Own your intelligence.