April 1, 2026
By Gautamdev Chowdary, co-founder and CTO, Zynix AI · Updated October 1, 2026
Multi-agent AI splits work among specialized agents coordinated by an orchestrator instead of asking one model to do everything. A Mount Sinai study in npj Health Systems found this design stayed accurate under heavy clinical workloads while using far fewer computing resources. For ACOs, the lesson is to judge AI by the workflows it completes, not the dashboards it produces.
Mount Sinai just published research that should reshape how every ACO leader thinks about their AI strategy. Their finding: orchestrated multi-agent AI systems outperform single-agent designs across clinical tasks — while using up to 65 times fewer computing resources.
This isn't an academic curiosity. It's a blueprint for how healthcare AI should actually work in value-based care.
For the past five years, the healthcare AI market has been dominated by monolithic platforms — single models that try to do everything from risk stratification to care gap closure to quality reporting. The result? Expensive implementations that generate dashboards nobody acts on.
The multi-agent approach flips this model. Instead of one system trying to be everything, you deploy specialized agents that each own a specific workflow, coordinated by an orchestrator that understands the broader care plan.
The timing of this research couldn't be more relevant. With 511 ACOs participating in MSSP in 2026, serving 12.6 million people with Traditional Medicare (a 12.3% increase from 2025), the LEAD model RFA released in March 2026, and ACO REACH ending at the end of 2026, organizations are making infrastructure decisions that will define their performance for years.
Here's a pattern we see repeatedly: an ACO invests in a population health platform, gets beautiful dashboards showing care gaps, risk scores, and quality measure performance. Six months later, the same gaps are still open.
Why? Because dashboards create visibility, not action. The care coordinator still has to manually review each patient, make the call, schedule the appointment, and document the outcome. The AI identified the problem. A human still has to execute the solution — at scale, across hundreds or thousands of patients.
Multi-agent systems solve this by distributing execution across specialized agents:
Each agent is optimized for its specific task. The orchestrator ensures they work together coherently. The result: care gaps get closed at machine speed with human oversight, not manual human execution.
CMS opened the LEAD model Request for Applications in March 2026. LEAD is a 10-year voluntary model running from 2027 through 2036 — the longest ACO commitment CMS has ever tested.
No benchmark rebasing over the full duration means that operational efficiency compounds. An ACO that automates care gap closure in year one doesn't just save money that year — it builds a structural advantage that grows every subsequent year.
For organizations evaluating LEAD, the infrastructure question is existential: can your current tech stack sustain performance improvement over a decade? If your care management runs on manual outreach lists, shared spreadsheets, and single-model AI that surfaces insights without executing workflows, the answer is almost certainly no.
At Zynix AI, we've been building agent-based infrastructure for ACOs since before the Mount Sinai research validated the architecture. Here's one example from a customer deployment:
Annual Wellness Visits are the scaffolding for ACO quality performance — they drive HCC capture, quality measure completion, and care plan updates. PBACO used our platform to move from manual AWV outreach to agent-driven patient engagement by voice and SMS across its provider network.
The data-platform category is moving the same way. Innovaccer, long known for unifying healthcare data, now pairs its platform with AI agents for clinical and administrative work.
The difference matters. Data platforms give you better visibility. Agent platforms give you better execution. In a world where every ACO has access to roughly the same claims data and risk models, execution is the differentiator.
Hippocratic AI is building patient-facing voice agents for non-diagnostic work such as pre-visit intake, scheduling and recovery check-ins. Abridge is expanding beyond ambient documentation into prior authorization. Epic is pushing AI Charting. Each is approaching the problem from a different starting point.
The gap in the market — and the opportunity for ACOs — is a unified agent-based platform that covers the full care continuum: identification, outreach, scheduling, documentation, and quality reporting, all coordinated by an orchestrator that understands your care model.
If you're evaluating your AI strategy — whether for LEAD, a new MSSP agreement, or simply improving current performance — here's the framework:
For every dashboard you have, ask: what happens after someone sees this data? If the answer involves a human manually executing a workflow that could be automated, you have an execution gap.
Instead of evaluating AI platforms by feature lists, evaluate them by workflows automated end-to-end. Can the system identify a care gap AND close it? Or does it just show you the gap?
If you're considering LEAD, model your infrastructure needs at year 5 and year 10, not year 1. Manual processes don't scale. Agent-based systems do.
AWV outreach, post-discharge follow-ups, preventive screening reminders — these are high-volume workflows where AI agents can show results early and build organizational confidence in the approach.
The Mount Sinai research confirmed what forward-thinking ACOs already know: the future of healthcare AI isn't one model trying to do everything. It's specialized agents working together, orchestrated for the specific workflows that drive value-based care performance.
The organizations that build this infrastructure now will have a decade-long head start.
It is a design in which several specialized AI agents each handle one task, such as finding patients with open care gaps, contacting them, booking visits or preparing documentation, while an orchestrator assigns work, manages priorities and hands exceptions to people. The alternative is a single model asked to do everything at once, which tends to lose accuracy as the workload grows.
Researchers at the Icahn School of Medicine at Mount Sinai compared one agent handling mixed clinical tasks with an orchestrator that gave each task to a dedicated worker agent. In the study, published in npj Health Systems in March 2026, multi-agent accuracy stayed high as batches grew from 5 to 80 tasks while single-agent accuracy collapsed, and the multi-agent runs used up to 65-fold fewer tokens.
Because a dashboard shows the gap but does not close it. Someone still has to review each patient, make the call, book the appointment and record the result, across hundreds or thousands of patients. When staff time is the constraint, gaps stay open however good the data is. Closing them takes execution capacity, which is where agents that run outreach and scheduling help.
Start with high-volume, rules-based work where the next step is clear: annual wellness visit outreach, post-discharge follow-up calls, preventive screening reminders and no-show recovery. These workflows have many patients, simple scripts and measurable results. Keep clinical judgment, complex conversations and care plan changes with your care team, and make sure agents route symptom questions to a clinician by rule.
The care team sets the rules: which patients agents contact, what they say, when they call and when to escalate. Agents handle outreach, scheduling, reminders and documentation prep, and route clinical questions, refusals and exceptions to staff. Licensed clinicians own clinical judgment. In the Zynix platform, care plans sequence agent and care team tasks so every step has an owner.
npj Health Systems: Orchestrated multi agents sustain accuracy under clinical-scale workloads compared to a single agent (Klang E, Omar M, Raut G, et al., 2026). https://www.nature.com/articles/s44401-026-00077-0
CMS: 2026 Medicare Accountable Care Organization Initiatives Participation Highlights (fact sheet). https://www.cms.gov/newsroom/fact-sheets/2026-medicare-accountable-care-organization-initiatives-participation-highlights
CMS Innovation Center: CMS Invites ACOs to Apply to the New LEAD Model. https://www.cms.gov/priorities/innovation/innovation-insight-cms-invites-acos-apply-new-lead-model
Abridge: Highmark Health and Abridge prior authorization collaboration (press release). https://www.abridge.com/press-release/highmark-health-ahn-abridge-prior-authorization
Epic: Epic AI Charting Rolls Out Alongside an Expanding Set of Built-in AI Capabilities. https://www.epic.com/epic/post/epic-ai-charting-rolls-out-alongside-an-expanding-set-of-built-in-ai-capabilities/
Hippocratic AI: company website. https://www.hippocraticai.com/
Innovaccer: company website. https://innovaccer.com/
CMS Innovation Center: LEAD (Long-term Enhanced ACO Design) Model. https://www.cms.gov/priorities/innovation/innovation-models/lead
Gautamdev Chowdary is co-founder and CTO of Zynix AI, where he leads engineering for the Zynix platform, its AI agents and ZynixLLM.