CrossnoKaye's AURA agents run read-only plant checks on AgentCore

Amazon Web Services published a post on 9 November 2026 in which CrossnoKaye's chief technology officer, Raphael Simon, and AWS solutions architect Anu Jayanthi describe AURA, an agent system for industrial plants. The authors say AURA runs on Amazon Bedrock AgentCore alongside CrossnoKaye's ATLAS control platform, which they say covers more than 100 facilities and 13,000 controlled assets. The post is vendor-published and the figures are the company's own.

What AURA does

AURA works in three modes. Chats are interactive investigations, which the post says can span up to 200 facilities in one turn. Tasks are repeatable investigations pinned to a fixed version. Routines are background watchers triggered by schedules, sensor thresholds, alarms or equipment mode changes.

The post's worked example is a night-shift catch on an industrial chiller plant. At 2:00 a.m. a Routine detects low-stage suction-pressure drift, runs its pinned Task in read-only mode, rules out sensor drift and low refrigerant charge, and isolates condenser-side fouling. Within four minutes of the trigger, an emailed brief with cited signal traces reaches the plant engineers. Before sending, AURA checks that every citation links to a saved telemetry window, setting-change record or sandbox output. The authors say no control change is made automatically.

How it keeps the numbers honest

The authors say early versions let the model write free-form Python that printed text, which was too brittle: a model would sometimes misname a column or drop part of a time window. Now the service stages up to 50 datasets in the sandbox and appends a verification harness to the model's script. The harness confirms each declared output is a table with the required columns and types, then saves it as Parquet. If the script fails or omits a column, the workflow turns the error into a retry prompt and re-runs in a clean sandbox before any conclusion is published.

Sandbox lifetime depends on the job. Interactive chats keep one warm session with a 15-minute idle timeout so an engineer's in-memory tables survive between questions. Scheduled Tasks and Routines get a fresh single-use sandbox per run, stopped when the calculation ends, to avoid cross-run contamination and idle charges.

The safety layers

• Guardrails and provenance: Amazon Bedrock Guardrails evaluates every prompt and completion; every brief needs verified citations and notes periods when sensor data was missing.
• Read-only sandboxes: the code sandboxes have no network route, credentials or API path to plant equipment, and a proposed setting change in chat needs explicit human approval under the operator's own identity.
• Edge base control: the plant's local cluster enforces state-machine guards, rate-of-change clamps, operating bounds and interlocks that, the authors say, cloud services, optimizers and AI agents cannot bypass.
• Hardware protection: safety relays, local trips and mechanical relief devices sit below the software.

What the company reports, and what it leaves out

CrossnoKaye reports a 34 percent cut in batch drying-time standard deviation, 6 percent more product within target yield, and more than 50,000 tons a year of recoverable production gap at quarries. It also reports over 5,000 AI agents monitoring more than 10,000 assets. These results are credited to the combined control, optimization and AURA stack, so the post does not isolate what the agents themselves contributed.

The post gives no model names, no cost for the models or AWS services, and no rate of wrong or missed diagnoses. Its one detailed diagnosis, a slipping belt that made a crusher fault look like a crusher problem, is a success. Citation checks prove a claim links to evidence; they do not prove the conclusion is right.

What a team can take from it

The transferable parts are controls, not products. Put numerical work in code with a checked output schema rather than in model prose. Run overnight watchers read-only. Keep hard limits in deterministic software where the model cannot reach them. Make missing data explicit: the authors note that a missing hour of sensor data is not the same as an hour with no events.