Crew 03: live, WaterDoctor, weekly cadence
Three agents read the week's research. One turns sensor data into a report a vet can read.
The crew runs inside WaterDoctor's backend every week. Three agents (research, fact-check and editor) sweep ten aquaculture journals, regional market feeds and government policy announcements, verify every item, and assemble a curated brief in English and 简体中文. A second pipeline reads each pond's sensor stream (pH, DO, temperature, ORP, ammonia, nitrite, nitrate, turbidity and algae), adds the week's research and the next seven days of weather, and produces a PDF that the farm manager, the vet and the environmental regulator can all read.
We built this for WaterDoctor. A research desk that learns the customer's domain, plus a report agent that turns operational data into auditable output a regulator can read, is the pattern we ship. This is one of six sample crews on this site.
The shape
Two pipelines, one weekly cadence and one human reviewer.
Deep research
Three agents (research, fact-check and editor) produce a curated, bilingual week-in-review of journals, regional market news and policy alerts. Each item is marked verified, unverified or flagged, and reviewer overrides stay on the record.
Weekly reports
For each pond, each week, the sensor stream becomes a core judgment, diagnoses, a disease checklist, an aeration strategy, a feeding schedule, FCR analysis and a cost breakdown, drawing on that week's research bundle and the next 7 days of weather.
Human gate
Every fact-check verdict can be overridden, and every weekly report passes a reviewer before it leaves the system. The crew does the legwork and an operator owns publishing.
The research desk
Research, then fact-check, then editor, so only verified items reach the report writer.
research.agent
Four parallel grounded searches
Two paper sweeps cover ten aquaculture journals: Aquaculture, Fish & Shellfish Immunology, Aquaculture Reports, Reviews in Aquaculture, Frontiers in Marine Science, Journal of Fish Diseases, Aquacultural Engineering, Marine Biotechnology, Nature Communications and Water Research. Two more cover market and policy news, tuned for CN or SEA. Every paper carries a real DOI, and every news item carries the publisher's exact date and URL, copied from the search grounding.
factcheck.agent
Verify, corroborate and gate
Each item is marked verified, unverified or flagged. An independent grounded search runs for each item and every candidate URL is HTTP-checked for liveness, keeping up to six live URLs per item. If more than 70% of items come back flagged on the first attempt, the batch retries once before the verdict is saved, so a bad search call can't set the result.
editor.agent
Bilingual cleanup
Tidies titles and harmonises EN and 简体中文 across every field, without inventing new metadata. Flagged items pass through untouched, so the reviewer sees exactly what fact-check saw.
Reviewer (human)
Overrides on the record
Per-item overrides and review notes attach to the run record. The report writer reads the same bundle the human signed off.
The weekly report
Sensor stream, research bundle and weather, in one bilingual PDF.
Each Monday, for each pond, the report writer pulls seven days of sensor readings (pH, dissolved oxygen, temperature, ORP, ammonia, nitrite, nitrate, turbidity and algae), the next seven days of weather for outdoor ponds, and the week's research bundle for the region. The result is a numbered report: overview, water quality (core judgment plus a sparkline and range per parameter), disease screening, weather risk and advice, aeration strategy, disease checklist, feeding schedule, FCR analysis, energy comparison and cost breakdown. Every section appears in EN and 简体中文 side by side. Every diagnosis links phenomenon to cause to remedy, with a HIGH, MEDIUM or LOW risk level a vet can defend.
What we won't ship
Five things the crew won't do.
No invented citations
The research agent's URLs are limited to those the search grounding returned. DOIs reconstructed from memory are dropped before fact-check sees them.
No date massaging
A paper outside the search window keeps its real date, and the report records it as it is.
No silent flags
Flagged items survive the editor untouched and appear in the reviewer panel, so the human sees exactly what the agent saw.
No publishing without a human
Every weekly report passes a reviewer before the EN and ZH PDFs go out. There is no auto-publish path.
No single-agent demo
A do-everything agent can't be evaluated. Five narrow agents with one job each are what's running in production.
A reusable pattern
We can build
the seventh for you.
Swap aquaculture journals for clinical guidelines, sensor streams for telemetry, and ponds for sites, and the shape holds: a research desk that learns your domain each week, a report agent that makes your operational data readable for your customer or regulator, and a human gate on every output. Our other sample crews cover business development, long-form articles, image production, product selection and finance. We can build the seventh for you.
Brief us