Agriculture AI agent · Livestock health
Livestock-Health Triage Agent (Herd & Flock Signals)
Watch sensor, production, feed, water and stockperson signals across the herd, rank the animals that need a look and escalate with their history — the veterinarian diagnoses, prescribes and sets any withdrawal period.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Diagnosis
- Vet only
- After launch
- Agent Care
What this agent does
Watches and escalates, does not diagnose
In
- 01
Take the signals as they arrive — collar, bolus, parlour, production, feed and water, and the treatment record.
- 02
Take what the stockperson saw on the walk-through, in the words they wrote, against the animal or pen it names.
Reason
- 03
Check the sensor itself first — off the animal, flat, unread or re-fitted to another animal is not a health signal.
- 04
Compare the animal against its own baseline and its pen-mates, so a group-wide shift is not read as one sick animal.
- 05
Check the day's management events — mixing, transport, a ration change, heat — before a drift is called disease.
Decide
- 06
Rank who needs a look first, with what moved, how long it has run and what else in the pen is doing the same.
- 07
Match the pattern against the client's reportable and welfare list and escalate it on the named path, ranking aside.
Out
- 08
Assemble the history a person needs — age, stage, prior events, treatments given and any withdrawal still running.
- 09
Retain the raw signals, the sensor's own status, the path taken and what the stockperson or the veterinarian found.
Product statement
The agent observes, ranks and escalates with the animal's history. Diagnosis, any treatment and any withdrawal period stay with a licensed veterinarian.
Example workflow
One animal, end to end
- 1Signals arriveCollar, bolus, parlour, production, feed and water data, and the stockperson's own note
- 2Sensor and identity checkedWhether the device is reading, on the animal it is recorded against, and current
- 3Baseline and pen comparedThe animal against its own history and against the pen-mates standing next to it
- 4History and events assembledAge, stage, prior events, treatments, any running withdrawal and the day's management log
No human action required
Stages 1 to 4 run without a person in the loop — the sensor check, the comparison and the history finish before anyone is asked to walk a pen. A reportable or welfare sign ends that stretch immediately, whatever the ranking says.
Enters the day's list in rank order.
Leaves on the named path at once, unranked.
Stockperson, then veterinarian
The stockperson walks the animal and records what they find; the veterinarian diagnoses, prescribes and sets any withdrawal period.
Go and look · Call the vet · Dismiss with a reason
Seen — handed back
- 6Case loggedWritten to the herd record only where write access and policy allow; the diagnosis field stays empty
- 7Outcome evaluatedDetection recall against confirmed cases, false-alert load, escalation-path correctness and record completeness
Corrections
A case the stockperson dismisses is recorded with the reason and counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Naming a disease or making a diagnosis.
- Prescribing, dispensing or directing a treatment.
- Setting, shortening or clearing a withdrawal period.
- Releasing an animal to slaughter, sale or the bulk tank.
Automation boundary
Agent acts unaided
- ✓Read sensor, production, feed, water and stockperson signals across the herd.
- ✓Compare each animal against its own baseline, its pen-mates and the day's events.
- ✓Assemble the animal's history, prior events, treatments and any running withdrawal.
- ✓Rank who needs a look and send a reportable or welfare sign out on the named path.
Write actions run only inside the approval boundaries agreed in implementation — never the diagnosis.
- Deciding a reportable sign need not be reported.
- Closing a welfare case without a person seeing the animal.
- Authorising euthanasia or a movement off the unit.
- Changing thresholds, escalation paths or the reportable list.
Example output
One animal on the list, annotated
Everything the agent returns is attached to the animal, the sensor it came from and the history around it.
Watch output · single animal, single morning
Illustrative example- Signal
- Rumination down, third day
- Animal
- Cow 4127, 12 days fresh
- Pen, time
- Pen 3, today 05:40
- Ranked
- First on this morning's list
- Confidence
- 88%
- Diagnosis
- The vet's, not the agent's
As received
The signal as the collar and the parlour recorded it, with the animal it is tied to and the pen she was standing in.
Evidence used
- Her own 14-day baseline
- Pen-mates unchanged
- Fresh-cow treatment record
Why she ranked first
Three days below her own baseline while the rest of the pen held steady.
What the score decides
Confidence sets her place on the list, not whether she is ill. It never holds an escalation.
Action
Go and lookCall the vetDismiss with a reasonValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every animal on the unit
Sensor, parlour, feed, water and the walk-through
Use the unit's own baselines
Compare against the animal's own history, the pen-mates beside it and the events your team logged that day — not a herd average from somewhere else.
Get the signal to a person the same day
A drift that would have surfaced at the next walk-through is on the morning list instead, with the history a person needs to judge it.
Send the serious ones out unranked
A reportable or welfare sign leaves on its named path the moment it is seen — ahead of the list and outside any confidence threshold.
Build an evidence trail
Retain the raw signal, the sensor's own status, the baselines compared, the events checked, the path taken and what the stockperson or the vet found — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Herd & treatment records
- Herd management systems
- DairyComp
- Treatment logs
- withdrawal records
Sensors & wearables
- Collars and ear tags
- rumination boluses
- Pedometers
- in-pen cameras
Production & parlour
- Parlour and milk meters
- conductivity
- SCC and milk testing
- growth records
Agent
Livestock-health triage
- Reads the signals
- Ranks the cases
- Escalates to a person
Feed, water & climate
- Feed and mixer systems
- water meters
- Barn climate sensors
- weather feeds
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the pen
Each control wraps the one inside it. A case clears every layer before it reaches a stockperson, and the diagnosis sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether the signal is real and whose it is
- L3
- Decides what leaves whatever the ranking says
- L4 – L5
- Keep diagnosis with a vet and the record intact
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the whole watch — not only the cases a stockperson walked.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the morning list the stockperson opens
- E1Final-output evaluationDid a case reach a person for the animals a vet later confirmed?
- E2Step-level evaluationDid it use the right baseline, the right pen and the day's event log?
- E3Tool evaluationDid it read the right device, the right animal and the right record?
- E4Escalation-path correctnessDid reportable and welfare signs leave on the named path, and how fast?
- E5Slice evaluationHow does detection change across housing systems, breeds and age groups?
- E6Business outcomeHow many alerts did a stockperson work per hour, and how many were nothing?
Floor — the animal a veterinarian ends up seeing
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle.
Agent lifecycle
Direction of processing →
01 · Signal intake
1 mode
- LH-01
Tag lost or on another animal
A missing or re-fitted device reads as healthy.
Stage gathers
The devices, the production data and the stockperson's note
02 · Comparison
2 modes
- LH-02
Group signal read as illness
A pen-wide cause is missed while animals are worked singly.
- LH-03
Management event alerted
Mixing, transport or a ration change fires as disease.
Stage tests
The animal against its own baseline, pen and the events
03 · Record / write
1 mode
- LH-04
Treatment record incomplete
The withdrawal period cannot be verified from it.
Stage writes
Only where write access and approval policy allow it
04 · Escalation
2 modes
- LH-05
Reportable sign not escalated
A notifiable pattern joins the ranked list instead.
- LH-06
Alert volume trains dismissal
So much fires that the list stops being walked.
Stage returns
The ranked list and the path a serious sign leaves on
05 · Change / Version
1 mode
- LH-07
Silent threshold regression
A tuning change moves the escalation line unnoticed.
Stage tracks
Model, prompt, threshold and sensor-feed changes
- Sev-1 · a sick animal or a duty is missed
- Sev-2 · the wrong picture reaches the vet
- Sev-3 · alerts degrade, the list is ignored
Affected slices
Detection is not spread evenly across the herd
A missed-case rate that reads acceptably across the whole unit can sit almost entirely in the animals a sensor fits worst and the groups the model has seen least of. Nestack reports performance by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Pre-weaning calves | 6.7% | 3.9× | Review | |
| Bedded-pack and loose housing | 4.8% | 2.8× | Review | |
| Sensor gaps and tag loss | 3.3% | 1.9× | Watch | |
| Mature cows, freestall housing | 1.4% | 0.8× | Normal |
Bar: missed-case rate lift vs. freestall-cow baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
A missed animal is answered in the pen before it is answered in the model
An animal found late cannot be found early again. The work is the vet call, the treatment record and the withdrawal — the loop decides who gets walked next.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Missed cases, false-alert load and escalation delay split by housing system and age group.
- 02
Diagnose
The device, the animal it was tied to, the baseline window and the event log are checked in turn.
- 03
Improve
The herd veterinarian agrees any change to a threshold, baseline or escalation rule, versioned.
- 04
Verify
Days already worked are scored again to see whether the missed animals now come up.
- 05
Learn
The animal's record becomes a test case, and the sign it showed goes on the standing watch list.
Learn → Detect
The return edge. A change to the reportable or welfare list is a veterinary change — your herd veterinarian signs it before it ships, not after.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, signals, comparison logic, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Sensor and herd-system API assessment.Week 1, Week 2
- 03Escalation paths, approved by your vet.Week 1, Week 2
- 04Animal identity, tag mapping and sensor status.Week 2
- 05Per-animal baselines and pen comparison.Week 2, Week 3
- 06Management-event log and confounders.Week 3
- 07Treatment, withdrawal and history assembly.Week 3, Week 4
- 08Ranked walk list and stockperson handoff.Week 4
- 09Detection-recall and alert-load evaluation.Week 4, Week 5
- 10Escalation-path testing on your own cases.Week 5
- 11Herd-system write-back and integration.Week 5, Week 6
- 12Observability, deployment and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne group and signal set | ProductionProduction herd-system integration | AdvancedMulti-site / multi-species |
|---|---|---|---|
| Introduced at Pilot | |||
| Signal watching and per-animal baselines | ✓ | ✓ | ✓ |
| Group versus individual separation | ✓ | ✓ | ✓ |
| Reportable and welfare escalation to a person | ✓ | ✓ | ✓ |
| Diagnosis left with a licensed veterinarian | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Management-event and confounder checks | — | ✓ | ✓ |
| Treatment and withdrawal history assembly | — | ✓ | ✓ |
| Ranked walk list and stockperson handoff | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-species and multi-housing coverage | — | — | ✓ |
| Multi-site and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on sensor and herd-system integrations, species and housing systems, herd or flock size, escalation paths, approval controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your sensors, how they are fitted and where they drop out
→Sensor and herd-system API assessment
Week 102Your reportable-disease and welfare escalation paths
→Escalation paths, approved by your vet
Week 103Animal identity, tag numbers and how they get re-used
→Animal identity, tag mapping and sensor status
Week 204Your event log — mixing, transport, ration and housing changes
→Management-event log and confounders
Week 305Your treatment records and how withdrawal is verified today
→Treatment, withdrawal and history assembly
Week 406Animals you found late, including the ones you nearly missed
→Evaluation suite, regression cases and failure-mode testing
Week 407Named stockpeople and your herd veterinarian to review
→Walk-list handoff, then supervised watching and validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 carries both the escalation testing and the first supervised watching.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Workflow discovery, escalation paths and the automation boundary
W2Animal identity, sensor status and the first signal set
W3Per-animal baselines, pen comparison and the event log
W4Detection-recall, false-alert-load and escalation-path testing
W5Herd-system integration, walk-list handoff and supervised watching
W6Stockpeople work from the live list, then Agent Care starts
- Reading the band
- Escalation-path correctness is tested in week 4, before a case reaches a stockperson in week 5. No case reaches the pen until it has.
- At the end of W6
- The list has run alongside your existing checks and been walked by the stockpeople who work from it, then Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Agriculture AI agent
Build a watching agent around your herd health routine.
Show us your sensors, your treatment records and how a sick animal reaches a stockperson and then the vet today. We'll walk the pens with your stockperson and agree what has to reach the vet the same day and what can wait for the next round.
- Typical delivery
- 4–6 weeks
- Deployment
- Your herd and sensor stack
- Diagnosis
- Licensed veterinarian, always
- After launch
- Agent Care available
Nestack Agents · Livestock-health triage
AGT-AGR-03 · Agent Care available after launch