Natera's AI Agent Cuts Instrument Failure Triage from 17.4 Hours to 8 Minutes
When a senior technician left Natera's lab, the knowledge of how to repair the company's instruments left with him. Engineers were taking 17.4 hours to answer a ticket in Jira. The SWARM agent, running on Amazon Bedrock AgentCore, reads each ticket itself, searches manuals and past tickets, and attaches a report with a ranked cause. The first response arrives in about 8 minutes, roughly 130 times faster, and 96.2% of tickets receive a report. The takeaway: if accuracy stalls, add documentation to the knowledge base first.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Imagine an instrument breaks down in the lab, and the only person who knew how to fix that equipment family has quit. Engineers open a Jira ticket, search for PDF manuals, and ask colleagues, and the first answer arrives 17.4 hours later.
That was the situation at Natera, known for its cell-free DNA (cfDNA) tests. For Natera, instrument downtime means delays in processing samples and patient reports. The lab engineering team, led by Ariel Zhirau, built the SWARM agent (Systematic Workflow Analytics and Resolutions Manager) with AWS on Amazon Bedrock AgentCore, AWS's managed service for running AI agents.
According to AWS, the system responds about 130 times faster (roughly 8 minutes instead of 17.4 hours) and produces an automated report for 96.2% of tickets.
What Problem Were They Solving?
Natera runs many instruments from different families, and each has its own typical failures, documentation, and standard repair methods. The service team depended on the personal experience of individual people rather than on unified instructions. When the expert who had spent years on one family left, his knowledge was effectively lost: there was no system to capture it and make it available.
The workload kept growing. The lab received more than 65 tickets a day across three Jira queues, and it could not hire engineers as fast as the volume grew. The goal was straightforward: respond within minutes, give accurate diagnoses across different equipment types, and keep improving by evaluating results.
How the System Works
Two agents work from a shared skill library. The swarm-agent picks up each new ticket before an engineer opens it, selects skills based on instrument type and symptoms, searches the knowledge base, and adds a link to an HTML report to the ticket. The report contains the leading hypothesis for the cause, ranked verification steps, and a confidence label. The swarm-hive handles dialogue: engineers clarify details, upload photos, and try alternative approaches, while AgentCore memory keeps the conversation context for each person. Because the dialogue accepts free text, requests pass through Bedrock Guardrails.
The library contains 33 skills: 24 for instrument families and 9 cross-cutting ones. Each skill has a clear interface and is tested separately, like ordinary production code. Python handles deterministic work, such as parsing Jira fields and formatting the report, while the LLM interprets symptoms and matches them to manuals. The practical benefit is clear: if a diagnosis is wrong, the team can point to a specific skill instead of debugging the entire system.
Tickets reach the agent through EventBridge Scheduler and a Lambda function that polls Jira, so no webhooks needed to be configured. From there, tickets go into an SQS queue, and DynamoDB removes duplicates by ticket ID and stores the status, cost, and token usage of each run. The queue decouples the agent from Jira's availability. X-Ray provides tracing, so a slow or expensive step is quick to locate.
The environment is clinical, so deployment uses no long-lived keys: GitLab CI receives temporary credentials through OIDC and AWS STS, traffic runs through private VPC subnets, and the remaining secrets are rotated every 30 days. The work runs on HIPAA-eligible AWS services under a BAA. Patient personal data does not enter either the vector index or the model: tickets contain only equipment and process metadata.
The Main Lesson: Data First, Model Second
Early prototypes that used standard search without equipment metadata gave unstable results, so the team wrote its own document-processing pipeline. Texts are split into chunks of 1,000 tokens with a 150-token overlap and indexed in batches of 100 documents using Titan Text Embeddings and Bedrock Knowledge Bases. For documents with tables and diagrams, the authors recommend a dedicated parser instead of standard chunking.
The most telling figure: after the team added missing vendor manuals for the weakest instrument family, accuracy rose from 71% to 75% with no changes to the model. According to Zhirau, diagnostic accuracy comes down primarily to data quality, so when performance plateaus, the team checks documentation coverage before it touches the model.
What to Take Into Your Own Project
- Don't hard-code logic into prompts: a versioned skill library makes auditing and debugging easier.
- Separate background batch work from interactive dialogue and run them on different runtimes.
- Hand everything deterministic to ordinary code and reserve the LLM for reasoning over text.
- Enable tracing and per-run cost accounting before the first release.
- When quality stalls, look for gaps in your documentation before you change the model.
Caveat: the figures come from AWS's own corporate blog, and the 71% and 75% apply to a single, weakest instrument family, not to the whole system. SWARM remains an assistant that prepares a suggested report, and the engineer makes the repair decision.
- Nov 21, 2026AG-UI 1.0 gives AI agents a common language for talking to apps: 31 event types
- Nov 21, 2026OpenAI launches ChatGPT Wallet: agents buy without ever seeing your card number
- Nov 21, 2026US Advertisers Spent $70.7 Million on ChatGPT Ads in October, Up 127% From April
- Nov 21, 2026ONNX Runtime vanishes from GitHub: Microsoft hides repositories after package hack
Key Questions Answered
What Problem Were They Solving?
Natera runs many instruments from different families, and each has its own typical failures, documentation, and standard repair methods. The service team depended on the personal experience of individual people rather than on unified instructions. When the expert who had spent years on one family left, his knowledge was effectively lost: there was no system to capture it and make it available. The workload kept growing.
How the System Works?
Two agents work from a shared skill library. The swarm-agent picks up each new ticket before an engineer opens it, selects skills based on instrument type and symptoms, searches the knowledge base, and adds a link to an HTML report to the ticket. The report contains the leading hypothesis for the cause, ranked verification steps, and a confidence label.
What are the key details on The Main Lesson: Data First, Model Second?
Early prototypes that used standard search without equipment metadata gave unstable results, so the team wrote its own document-processing pipeline. Texts are split into chunks of 1,000 tokens with a 150-token overlap and indexed in batches of 100 documents using Titan Text Embeddings and Bedrock Knowledge Bases. For documents with tables and diagrams, the authors recommend a dedicated parser instead of standard chunking.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.