How Cohere Well being digitizes scientific insurance policies utilizing Amazon Bedrock AgentCore

Prior authorization is the approval course of well being plans require earlier than masking sure medical companies or drugs. It stays some of the guide processes in healthcare, not as a result of the medical reasoning for requiring approval is flawed, however as a result of the insurance policies that govern it are trapped in static, unstructured codecs that resist automation. This content material is on the core of day-to-day scientific operations impacting a whole bunch of tens of millions of sufferers annually. Nevertheless, the coverage content material varies by scientific space, geography, line of enterprise, and well being plan, and evolves as medication and expertise advances. Traditionally, well being plans didn’t have a scientific solution to handle, analyze, and optimize them. Digitizing these scientific insurance policies into structured, machine-readable information utilizing customary terminologies scale back a important operational bottleneck by supporting extra constant, computable workflows and serving to well being plans modernize prior authorization operations at scale whereas sustaining applicable scientific oversight.

Cohere Health^(R), a scientific intelligence firm that powers well being plan operations, constructed Cohere Policy Studio^(TM) utilizing Amazon Bedrock AgentCore, which gives the multi-tenant isolation required for his or her well being plan clients and a managed agent runtime that accelerates deployment with out rebuilding infrastructure. The appliance makes use of a versatile, multi-tenant agentic structure to speed up coverage digitization with in depth workflow administration and computerized model monitoring.

On this put up, you find out how Cohere Well being constructed a multi-tenant agentic structure on AgentCore utilizing AgentCore Runtime’s safe MicroVM isolation, unified device entry by way of AgentCore Gateway, AgentCore Reminiscence, and the Agent Skills open customary to quickly scale coverage digitization capabilities, whereas preserving transparency, model management, and human oversight.

Problem: The coverage digitization bottleneck

Realizing the worth of AI-assisted workflows in prior authorization depends upon a foundational problem: reworking the principles trapped in static paperwork and PDFs into structured, machine-readable information that AI methods can use extra persistently, whereas medical skilled stay liable for scientific assessment the place scientific judgment is required. Well being plans face a posh problem of managing scientific insurance policies to help quickly altering necessities. Automating coverage digitization helps well being plans adapt to those modifications.

Cohere Well being recognized three challenges in constructing an AI resolution for this workflow:

• Authorities rules – Per Facilities for Medicare & Medicaid Companies (CMS) regulations, well being plans are required to help API-based digital prior authorization by January 2027.
• America’s Well being Insurance coverage Plans (AHIP) – The AHIP commitments require well being plans to realize 80 p.c real-time approvals for digital prior authorization submissions. Every line of enterprise has distinctive necessities, growing the necessity to rapidly handle, audit, and deploy scientific insurance policies.
• Technical structure calls for – The answer wanted to ingest a number of enter codecs and produce completely different representations of every coverage for various downstream shoppers, every with its personal suggestions loop.

AgentCore addresses these challenges with managed runtime infrastructure, session isolation, and unified device entry.

Answer overview

The next diagram exhibits how Cohere Coverage Studio connects AgentCore Runtime, Gateway, and Reminiscence right into a unified agentic system for coverage digitization.

The Coverage Studio software is constructed on AgentCore utilizing the Agent Abilities open customary. To scale out representations in Cohere Coverage Studio, Cohere Well being added new expertise to an present AgentCore Runtime that was already decomposing insurance policies. This runtime had entry to the coverage expertise, coverage APIs as Model Context Protocol (MCP) instruments by way of AgentCore Gateway, and session reminiscence for coverage analysts’ suggestions loops, serving to groups refine outputs inside a ruled, human-in-the-loop course of.

The staff accomplished three duties:

• Deployed AgentCore Runtime with AgentCore Gateway and AgentCore Reminiscence for a full agentic system utilizing LangChain.
• Configured AgentCore Gateway to fetch instruments and expertise.
• Wrote expertise with scientific coverage consultants and evaluated them utilizing Cohere Well being’s standardized observability course of primarily based on Arize AI.

You possibly can apply these identical patterns to construct your individual multi-tenant agentic system.

Deploying AI brokers with reusable Amazon Elastic Container Registry (Amazon ECR) base photos

Cohere Well being serves a number of well being plans that require strict information isolation between tenants. AgentCore Runtime’s safe microVM isolation enforces this with devoted compute, reminiscence, and filesystem assets per session.

When deploying a number of AI agent situations throughout groups, sustaining consistency whereas permitting customization is necessary. Every staff wants its personal agent configuration, however rebuilding the whole runtime setting for each deployment creates pointless overhead and drift. You need to use the next base picture sample to deploy new brokers to AgentCore Runtime microVMs with a minimal Dockerfile.

Base picture and shopper sample

Cohere Well being developed a two-tier deployment structure that separates the secure runtime setting from team-specific configurations:

FROM {account_id}.dkr.ecr.{aws_region}.amazonaws.com/cohere-agent:v1
COPY agent_config.yaml /app/src/agent_config.yaml

The FROM line pulls the shared base picture containing the LangChain agent framework and customary dependencies. The COPY line provides the team-specific agent_config.yaml, which controls the next choices:

• Reminiscence modes – Select between stateless (NO_MEMORY) or persistent (AGENTCORE) dialog historical past.
• Storage methods – full_trace for correction workflows or conversation_only for clear historical past.
• Session context caching – Routinely caches talent definitions and paperwork to keep away from redundant Amazon Easy Storage Service (Amazon S3) fetches.
• Immediate caching – Might help scale back prices and latency by caching system prompts and incessantly used content material.
• Versatile device configuration – Allow/disable instruments per deployment.
• Mannequin configuration – Base mannequin on Amazon Bedrock with configurable token limits, temperature, and different inference parameters.
• LiteLLM configuration – Configure LiteLLM because the reverse proxy between the mannequin and the agent.

With the runtime deployed, the following step was connecting it to instruments and expertise.

Unified device and talent entry with AgentCore Gateway

Cohere Well being’s brokers entry a number of device varieties, together with AWS Lambda features for fetching expertise and paperwork, and inside APIs, maintained throughout completely different groups. AgentCore Gateway consolidates these behind a single authenticated endpoint, so groups add new instruments with out redeploying the agent.

AgentCore Gateway structure

Cohere Well being applied this utilizing AgentCore Gateway with separate targets for shared instruments and project-specific instruments.

Instrument Lambda operate construction

AgentCore Gateway invokes an AWS Lambda operate for every device request. The operate routes to the proper handler primarily based on the device title handed within the gateway context.

# jobs/generic-tools-lambda/app.py
import json
from instruments.fetch_skill import handler as fetch_skill_handler

# Routing dictionary for device discovery
TOOL_HANDLERS = {
    "fetch_skill": fetch_skill_handler
}

def lambda_handler(occasion, context):
    """Gateway-compliant Lambda handler with MCP routing"""

    # Extract device title from gateway context
    tool_name = context.client_context.customized.get('bedrockAgentCoreToolName', '')

    # Strip gateway prefix (gateway provides {goal}__ to device names)
    if '__' in tool_name:
        tool_name = tool_name.cut up('__', 1)[1]

    # Path to applicable handler
    if tool_name not in TOOL_HANDLERS:
        return {
            "statusCode": 404,
            "physique": json.dumps({"error": f"Instrument {tool_name} not discovered"})
        }

    attempt:
        end result = TOOL_HANDLERS[tool_name](occasion)
        return {
            "statusCode": 200,
            "physique": json.dumps(end result)
        }
    besides Exception as e:
        return {
            "statusCode": 500,
            "physique": json.dumps({"error": str(e)})
        }

Instrument implementation

Every device handler fetches information from a selected supply. The next instance retrieves a talent definition from Amazon S3.

# instruments/fetch_skill.py
import boto3
import os

def handler(occasion: dict) -> dict:
    """Fetch talent definition from S3"""

    skill_id = occasion.get('skill_id')
    if not skill_id:
        return {"error": "skill_id required"}

    # Use setting variables for configuration
    bucket = os.environ.get('SKILLS_BUCKET')
    prefix = os.environ.get('SKILLS_PREFIX')

    s3 = boto3.shopper('s3')

    attempt:
        response = s3.get_object(
            Bucket=bucket,
            Key=f"{prefix}/{skill_id}.yaml"
        )
        content material = response['Body'].learn().decode('utf-8')

        return {"content material": content material}
    besides Exception as e:
        return {"error": f"Didn't fetch talent: {str(e)}"}

Agent configuration

The agent configuration defines which gateway targets the agent can entry and the way it authenticates.

# agent_config.yaml
mcp:
  gateway_url: {gateway_url}
  allowed_targets:
    - generic-tools    # AIP-maintained instruments
    - digitization-tools  # Undertaking-specific instruments
  auth_mode: "bearer_token"

With the runtime and instruments in place, Cohere Well being turned to constructing the area experience layer.

Abilities improvement and analysis

AI brokers want domain-specific data to carry out specialised duties successfully. Generic prompts produce inconsistent outcomes, require in depth token utilization, and lack the nuanced understanding that area consultants convey. Every new use case historically required rebuilding agent infrastructure from scratch, creating bottlenecks in deployment velocity. A modular expertise framework addresses this by decoupling area experience from infrastructure. For Cohere Well being, this implies scientific coverage consultants can writer and refine new expertise immediately, serving to make sure the system helps coverage workflows in ways in which stay grounded in professional assessment and governance.

Modular expertise framework

Groups deploy new capabilities by way of modular, versioned talent definitions with out rebuilding the agent.

Growth workflow

Cohere Well being follows a structured workflow to develop and validate every talent earlier than it reaches manufacturing.

Analysis course of

Evaluating expertise requires collaboration between machine studying engineering and information science. The method begins with reference datasets that include floor reality outputs for every talent. The staff defines success metrics (accuracy, completeness, and consistency) and runs an analysis suite in opposition to these take a look at instances. When a talent fails, the staff analyzes the failure mode and iterates on the talent definition earlier than retesting.

After a talent passes the analysis suite, information science evaluations the outcomes in opposition to acceptance standards and approves the talent for manufacturing deployment.

After deployment, Arize AI tracks effectiveness metrics in manufacturing. Medical coverage analysts annotate pattern outputs to catch errors the automated metrics miss. The staff screens for talent degradation over time and makes use of these information factors to prioritize optimization work.

Talent versioning and deployment

Abilities transfer to manufacturing by way of a layered versioning scheme and a staged deployment pipeline.

Twin-layer versioning

Abilities use dual-layer versioning: semantic versioning for functionality monitoring and Amazon S3 object versioning for deployment historical past. The primary layer tracks functionality modifications in SKILL.md, with every model tagged in git (for instance, talent/policy_ingestion/v1.2.3). Amazon S3 object versioning gives the second layer, sustaining immutable historical past for each add with rollback functionality and separate non-prod/prod buckets.

Deployment circulation

1. Developer commits and opens a PR to develop.
2. Steady integration and steady supply (CI/CD) packages talent.tar.gz with metadata on merge.
3. The pipeline uploads to the Amazon S3 non-prod bucket and updates the manifest.
4. Consider in non-prod setting.
5. Open PR to most important.
6. Deploy to prod with gradual rollout and monitoring.

Outcomes and influence

By this implementation, Cohere Well being achieved measurable enhancements throughout coverage digitization velocity, deployment velocity, and protection.

Coverage digitization effectivity: Total time spent on coverage digitization decreased by 30 p.c, from 2 hours quarter-hour to 1 hour 35 minutes per coverage. Cohere Well being has digitized hundreds of insurance policies thus far utilizing guide and semi-automated workflows. The agent-based framework targets additional time discount per coverage because it scales throughout the prevailing coverage library.

Deployment velocity: Full agent deployments within the product decreased from 3–4 months to 2–6 weeks. The reusable ECR base picture sample lets groups get up a brand new agent with a minimal Dockerfile, and the modular expertise framework means new capabilities ship with out rebuilding the agent runtime. The system abstracts DevOps considerations, so conventional machine studying (ML) and information science engineers can deploy brokers with out in depth coding expertise. The coverage digitization product runs a single-agent, multi-skill structure with one agent, a major talent with a sub-skill, and three reference injections.

Coverage protection: Cohere Coverage Studio represents coverage content material with verbatim textual content and a normal codified proof layer, packaged collectively and accessible throughout authentic coverage codecs and sources.

“Prior authorization coverage assessment has at all times demanded a rare degree of scientific consideration—each phrase in a coverage doc can carry downstream penalties for sufferers. However that spotlight has traditionally been cut up between interpretation and verification: not simply understanding what a coverage means clinically, however confirming which model of it ruled a given choice, and whether or not that very same model is what the well being plan revealed to suppliers. These aren’t administrative questions—they’re questions that bear immediately on scientific integrity. Amazon Bedrock AgentCore gave us the structure to handle each concurrently—AI-powered agentic workflows that help with navigating the interpretive complexity of scientific language, with built-in reminiscence and model monitoring that make provenance a first-class concern fairly than an afterthought. Structured, versioned coverage outputs make the scientific foundation of a choice traceable and reviewable by design, and AgentCore’s safe, multi-tenant runtime means we will ship that functionality throughout each well being plan we serve with out compromising isolation.”

— Brian Covino, M.D., FAAOS, Chief Medical Officer, Cohere Well being

Apply these patterns to realize comparable outcomes: reusable base photos for constant deployments, unified device entry by way of a single gateway, and modular expertise that scale with out rebuilding infrastructure.

Future: Connecting insurance policies by way of a data graph

Constructing on Cohere Coverage Studio’s success with AgentCore, the following evolution introduces an clever data graph which is already underway. Working with the AWS Generative AI Innovation Heart, Cohere Well being prototyped the foundational semantic layer mapping scientific insurance policies to standardized ontologies (UMLS, SNOMED) to help higher interoperability utilizing standardized healthcare phrases. Utilizing Amazon Neptune, this grounds coverage ideas in a construction that AI can traverse and hint. That graph connects scientific insurance policies with decisioning merchandise throughout expanded indications.

Enhanced structure

The data graph layer sits between the coverage illustration engine and downstream decisioning methods, making a semantic community that:

• Maps relationships between insurance policies, scientific tips, medical codes (ICD-10, CPT, HCPCS), drug formularies, and prior authorization standards throughout therapeutic areas.
• Scales indication protection by figuring out patterns and similarities throughout scientific domains, in order that new coverage varieties deploy quickly with out guide configuration.
• Connects coverage fragments to a number of decisioning contexts, so {that a} single coverage replace propagates appropriately throughout affected authorization workflows.

Key capabilities

As new insurance policies are digitized by way of AgentCore, the data graph is designed to assist determine related connections, flag potential conflicts, and recommend reusable patterns to help reviewer and coverage staff workflows. The graph learns from coverage constructions throughout scientific areas, suggesting templates and accelerating time-to-deployment for brand spanking new indication varieties from days to hours. Decisioning engines question the data graph utilizing pure language or Quick Healthcare Interoperability Assets (FHIR) assets to retrieve doubtlessly related coverage fragments with full provenance and model historical past. The graph additionally maintains bidirectional hyperlinks between CMS necessities, AHIP commitments, and inside coverage representations, supporting regulatory alignment at scale.

These capabilities ship complete indication protection with out proportional engineering effort, real-time coverage updates throughout linked decisioning merchandise, automated battle detection to assist forestall inconsistent authorization outcomes, and sub-second coverage retrieval for authorization requests.

This data graph basis helps Cohere Well being’s skill to assist well being plans obtain 80 p.c of digital prior authorization approvals in actual time. The graph maintains the safety, multi-tenancy, and audit capabilities established within the present AgentCore structure.

Conclusion

On this put up, you discovered how Cohere Well being used AgentCore and three architectural choices to cut back AI agent deployment from months to weeks. Three patterns (reusable ECR base photos, unified device entry by way of AgentCore Gateway, and modular expertise improvement) helped Cohere Well being help extra scalable coverage digitization workflows throughout codecs whereas decreasing digitization time by 30%.

The ECR base picture sample alleviates redundant infrastructure work, so groups can deploy new brokers with a minimal Dockerfile. Cohere Well being can scale the AI system with out rebuilding the runtime. The AgentCore Gateway structure gives a single authenticated endpoint for the instruments, whether or not they’re utilities primarily based on AWS Lambda or OpenAPI companies. The talents framework, constructed on the Agent Abilities open customary, separates area experience from agent mechanics, supporting fast iteration with steady analysis by way of Arize AI and scientific coverage analysts.

The way forward for healthcare AI depends upon methods that may adapt rapidly to altering necessities whereas sustaining reliability and safety. With AgentCore and these architectural patterns, you’ll be able to construct that system right now.

To get began with these patterns in your individual setting, discover the next assets:

Study Cohere Well being’s different AgentCore deployment of a medical necessity assessment agentic assistant on this re:Invent session.

In the event you’re a startup constructing production-ready AI brokers, AWS Activate gives the credit, technical steerage, and structure help that will help you transfer from prototype to manufacturing. Get began right now.

In case you have suggestions or questions on this put up, depart a remark within the feedback part.

Concerning the authors