Natera builds agentic AI to automate lab equipment diagnostics

The problem of lost expertise

Natera, a company specializing in cell-free DNA testing, faces a significant challenge in maintaining its laboratory infrastructure. The company relies on a wide variety of precision instruments, each with unique failure modes and complex vendor documentation. Because the engineering team relies on deep individual expertise rather than standardized playbooks, the departure of a senior technician can result in the loss of years of specialized knowledge.

This lack of centralized knowledge creates a bottleneck for lab operations. With over 65 tickets arriving daily across three different queues, the average first response time was 17.4 hours. This delay directly impacts the processing of patient samples and the speed of reporting results. Hiring more engineers was not a viable solution to keep pace with the volume of incoming support requests.

To solve this, Natera built SWARM, a Systematic Workflow Analytics and Resolutions Manager. The goal was to create an agentic AI system that could capture and deploy institutional knowledge at scale. By using AI to handle the initial diagnostic heavy lifting, the company aims to provide engineers with immediate answers the moment a ticket is submitted.

Building on Amazon Bedrock

Natera chose to build SWARM using Amazon Bedrock AgentCore, a managed service designed to build and operate AI agents. The team needed a system that could handle high-volume ticket processing autonomously while integrating with existing tools like Jira and Confluence. They specifically wanted a managed runtime to avoid the burden of building and maintaining custom infrastructure for security and scaling.

A key requirement for the project was a secure environment suitable for a clinical setting. The team implemented a keyless deployment model using OpenID Connect and AWS Security Token Service to generate temporary credentials. This ensures that no long-lived keys are stored in the pipeline, which is critical for maintaining the security of a regulated laboratory environment.

The architecture also prioritizes observability to ensure the system remains reliable. Because agentic systems can be difficult to debug, Natera integrated tools like AWS X-Ray to trace every run from intake to completion. This allows the team to pinpoint exactly which part of the AI's reasoning or which specific tool failed during a diagnosis.

A library of specialized skills

Rather than embedding all logic directly into the AI's prompts, Natera opted to build a separate, versioned skill library. This architectural choice makes the system easier to audit and improve because engineers can point to a specific skill if a diagnosis is incorrect. The library currently contains 33 skills, including 24 for specific instrument families and 9 for broader workflow tasks.

The system splits these skills into two categories: deterministic tasks and language reasoning. Python handles the deterministic work, such as parsing fields from Jira or formatting HTML reports. Meanwhile, large language models handle the reasoning tasks, such as interpreting symptoms, matching them against vendor manuals, and generating a narrative for the diagnosis.

This distinction is a high-impact design choice for teams building multi-skill agents. By treating these skills as production software rather than just prompt instructions, Natera can test and evaluate each component independently. This structure allows the system to grow as new equipment is added to the lab without requiring a total rewrite of the agent logic.

Knowledge and data quality

The accuracy of SWARM depends heavily on the quality of the data it can access. The Natera team found that general-purpose retrieval produced inconsistent results, so they built a custom document processing pipeline. This pipeline is specifically tailored to handle technical equipment documentation, including complex tables and diagrams that standard chunking methods often struggle with.

The team uses Amazon Titan Text Embeddings to create vector representations for semantic search. They discovered that diagnostic accuracy is often a data quality problem rather than a model problem. For example, adding missing vendor manuals for a low-scoring instrument family lifted accuracy from 71% to 75% without any changes to the underlying AI model.

To maintain the system, the team uses a continuous evaluation loop. Every time the agent processes a ticket, it provides a lead cause hypothesis and a self-rated confidence label. Engineers then grade these responses as Accurate, Partial, or Wrong. This feedback loop allows the team to identify gaps in the knowledge base and refresh the vector index without disrupting active lab operations.

Two runtimes for two needs

Natera deployed two distinct runtimes to handle different types of work within the SWARM system. The first, called swarm-agent, is designed for high-throughput background analysis. It automatically selects skills based on equipment types and symptoms to produce a structured HTML report with a confidence label, which is then linked to the Jira ticket.

The second runtime, swarm-hive, is designed for interactive collaboration between engineers. This runtime allows humans to engage in follow-up conversations, upload images, and explore different troubleshooting paths in real time. It uses AgentCore memory to retain the conversational context for each specific engineer across different sessions.

By separating these two runtimes, Natera ensures that the system meets different latency and context requirements. The swarm-agent provides rapid, automated summaries for the majority of tickets, while swarm-hive provides a deep-dive environment for complex issues. This dual approach allows the system to scale efficiently while still supporting the nuanced needs of human experts.