Overview of the Medera Agentic Framework

The Medera Agentic Framework is a modular AI system for product teams building clinical and operational behavioral-health workflows — without spending months on agent architecture. It ships with a canvas-based workflow engine (16 core nodes + 8 BH-OS rails), a multi-agent care orchestrator, and a typed expert registry. It is designed to support use cases across behavioral health — from patient-facing intake voice agents to clinician-facing scribes, multimodal co-therapy, and back-office prior authorization.

---

​

What problems it solves

1. LLMs do not have reliable access to clinical data LLMs cannot be trusted to rely on internal knowledge alone. In behavioral health, responses must be grounded in validated screeners, payer policy, chart context, and peer-reviewed evidence. Without access to these sources at runtime, models are forced to infer or guess — which is unacceptable in clinical settings. The Medera Agentic Framework gives agents direct access to a typed expert registry (PHQ-9, C-SSRS, ICD-10, PubMed, DrugBank, Payer Policy, …). Instead of hallucinating, agents look things up, verify context, and base their outputs on authoritative sources.

​

2. LLMs cannot safely act on the world

Clinical workflows involve interacting with real systems — querying EHRs, drafting documentation, generating referrals, submitting prior authorizations, and triggering downstream processes. The framework provides a controlled execution layer that lets agents plan actions, invoke experts, and coordinate multi-step workflows within explicit safety boundaries. Where necessary, agents pause execution at a user-approval node, request a human decision, and resume only after explicit consent is recorded in workflow_approvals.

---

​

What you can build

---

​

Built for behavioral health by design

Behavioral health is not a general-purpose domain, and the platform reflects that reality.

​

Safety first

Typed inputs and outputs, explicit expert schemas, guardrail nodes for HIPAA / 42 CFR Part 2 / suicide-language detection, and a non-bypassable crisis escalation path.

​

Auditability

Every decision, expert invocation, and tool call is recorded in execution_trace[]. Deployment actions write to a WORM deployment_audit_log chained with Merkle integrity checkpoints.

​

Domain-specific reasoning

gpt-4o for SOAP and post-visit analysis (T=0.3, JSON mode), gpt-4o-mini for real-time NLP, Medera reasoning for voice agent LLMs (T=0.3, max 500 tokens). Tuned for behavioral-health language.

​

Multi-agent architecture

Four voice agents and six workflow agents coordinated by the Care Orchestrator with an event bus carrying typed events (INTAKE_*, PA_*, REFERRAL_*, …).

​

Memory and context management

Per-execution with five variable scopes (workflow.input, state.*, nodes.*.output, runtime.context, tenant.secret) plus persistent agent memory under thread, patient, or tenant scope.

​

Prebuilt experts

A typed library of clinical capabilities — Vocal Prosody, Linguistic Content, Validated Screener, C-SSRS Risk, ESI Acuity, ICD-10, CPT / E&M, PubMed, DrugBank, ClinicalTrials.gov, Payer Policy, and more.

​

Third-party integrations

Plug directly into Epic, Cerner, athenahealth, Allscripts, eClinicalWorks, NextGen via FHIR R4; e-Rx network for e-prescribing; billing clearinghouse for billing; and MCP servers for custom tools.

​

Runtime context

Pass relevant context per request — FHIR resources, screener scores, patient demographics — so experts work with rich, domain-specific information.

---

​

Who it’s for

The Medera Agentic Framework is built for teams working on behavioral and mental health software: - Healthcare software companies embedding intelligent automation directly into their products

• Health systems building internal AI-powered clinical workflows
• Advanced engineering teams that need flexibility, control, and safety guarantees without building bespoke agent infrastructure The platform is not limited to prompt-based chatbots. It is designed to go from demo to production-grade clinical AI systems that operate safely in real-world settings.

---

​

Agents vs. workflows

Understanding the difference helps you pick the right approach. Agents are autonomous systems that reason and adapt. They use the LLM to understand context, make decisions, and take actions even in scenarios they haven’t seen before. Agents excel at open-ended tasks that require judgment — intake, triage, multimodal analysis, and PA appeals. Workflows are structured, step-by-step processes that follow predefined paths. They execute in a fixed order — like a recipe or a checklist. Workflows are ideal for repeatable processes that demand consistency and compliance — measurement-based care, eligibility verification, claims scrubbing. In Medera, the Canvas lets you blend both — agent nodes for judgment-heavy work, deterministic logic nodes (condition, transform, classify, wait, user-approval) for compliance-heavy work, and guardrail nodes that span both.