Agriculture AI agent · Crop scouting
Crop-Scouting Computer-Vision Agent (Flag & Locate)
Read drone, satellite and ground imagery for anomalies, geolocate each one and rank the walk — the cause is confirmed in the field, and any treatment belongs to an agronomist and a certified applicator.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Cause confirmed
- In the field
- After launch
- Agent Care
What this agent does
Finds and locates, does not diagnose
In
- 01
Take imagery from the flight, the satellite pass, a ground rig or a phone, with its date, sensor and capture conditions.
- 02
Take the field boundary, the crop, the variety, the planting date and the operations already recorded for that field.
Reason
- 03
Flag anomalies — weed pressure, lesions, discolouration, stand gaps, lodging and water stress — and size each one.
- 04
Geolocate every flag to a point and a zone a scout can find, not to a coloured area of a map.
- 05
Compare with the previous capture of the same field and separate crop change from capture change.
Decide
- 06
Mark imagery whose light, cloud, blur or angle makes the read unreliable, instead of scoring it as though it were clean.
- 07
Route a degraded capture, or a crop, cultivar or stage the model saw little of, to the agronomist before anyone walks.
Out
- 08
Produce a scouting report — what, where, how large, how certain, and the look-alikes the scout has to rule out.
- 09
Retain the imagery, the conditions, the model version and what the scout found when they got to the flag.
Product statement
The agent flags and locates. What the symptom is, and whether anything is applied, stay with an agronomist and a certified applicator.
Example workflow
One capture, end to end
- 1Imagery receivedDrone flight, satellite pass, ground rig or a phone photo, with the sensor, altitude and conditions
- 2Field context gatheredBoundary, crop, variety, planting date, growth stage and the operations already on record
- 3Anomalies flagged and locatedWeed pressure, lesions, discolouration, stand gaps, lodging and water stress, each placed and sized
- 4Change and capture quality checkedCompared with the last capture of the same field, with sensor, sun-angle and cloud differences separated out
No human action required
Stages 1 to 4 run without a person in the loop — flagging, locating and the change check finish before anyone is asked to walk anything. A capture the agent cannot read is returned as unreadable, not scored.
Enters the walking order with its look-alikes.
Goes to the agronomist before anyone walks.
Agronomist or scout
Walks to the located flag, splits stems, digs roots or pulls a tissue sample where that is what settles it, and records the cause.
Confirm · Reclassify · Send to the lab
Walked — handed back
- 6Scouting report issuedWritten to the farm record only where write access and policy allow; no product, rate or prescription is created
- 7Outcome evaluatedAgreement with the walk, look-alike confusion, geolocation error and misses by crop, cultivar and stage
Corrections
A flag the scout reclassifies on the ground is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Naming the cause of a symptom as confirmed.
- Turning a flag into a product, rate or prescription.
- Deciding an economic threshold has been crossed.
- Declaring a field clear of a pest or disease.
Automation boundary
Agent acts unaided
- ✓Flag anomalies in the imagery and geolocate each one to a findable point.
- ✓Compare with the previous capture and separate crop change from capture change.
- ✓Mark a capture unreadable and say what made it unreadable.
- ✓Rank the flags into a walking order with the look-alikes named against each.
Write actions run only inside the approval boundaries agreed in implementation. The report is a walking list, not a plan.
- Sending a scout into a restricted-entry interval.
- Closing a flag without anyone walking it.
- Reporting a regulated or notifiable pest onward.
- Changing capture rules, class lists or thresholds.
Example output
One flag, annotated
Everything the agent returns is attached to the capture it came from and the place it points to.
Scouting output · single flagged anomaly
Illustrative example- Capture
- Drone flight, overcast, midday
- Crop and stage
- Soybean, beginning pod
- Field, date
- Field 12, this morning
- What was seen
- Interveinal chlorosis, patchy
- Confidence
- 88%
- Cause
- Settled on the walk, not here
As received
The imagery as captured, with the sensor, the altitude and the light at the time.
Evidence used
- Previous capture, same field
- Growth-stage record
- Pattern and extent
Why this is not a diagnosis
Brown stem rot, sudden death syndrome and fungicide injury all look like this from above.
What the score decides
Confidence decides where this sits in the walking order, not what the problem is.
Action
ConfirmReclassifySend to the labValue
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 capture
From the flight, the satellite pass or the ground rig
Apply the field's own record
Use the boundary, the crop, the variety, the planting date, the growth stage and the operations already recorded for that field.
Start the walk where something is
Anomalies are found, sized and located before anyone leaves the truck, so scouting time goes to the acres that have something on them.
Send the doubtful reads to a person
A degraded capture, an unfamiliar cultivar or a symptom whose look-alikes the imagery cannot separate reaches the agronomist before it reaches a walking list.
Build an evidence trail
Retain the imagery, the capture conditions, the model version, the flag and its location, and what the scout found on the ground — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Imagery sources
- Drone flights
- satellite passes
- Ground rigs
- phone photos
Farm records & boundaries
- Operations Center
- FieldView
- Trimble Ag
- field boundaries
Agronomy & scouting
- Scouting apps
- agronomy platforms
- Diagnostic labs
- co-op systems
Agent
Crop-scouting computer vision
- Reads the imagery
- Flags and locates
- Ranks the walk
Weather & conditions
- Forecast
- in-field stations
- Growing-degree and rainfall records
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 walking list
Each control wraps the one inside it. A flag clears every layer before a scout is sent to it, and the cause and the treatment sit outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether the read may stand
- L3
- Decides when the model is outside what it knows
- L4 – L5
- Keep the cause with a person and the record intact
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the whole read — not only the flags on the map.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the scouting report the agronomist opens
- E1Final-output evaluationDid the flags agree with what the scout found on the ground?
- E2Step-level evaluationWas the symptom class right, or was a look-alike returned instead?
- E3Tool evaluationDid it read the right field, the right capture and the right boundary?
- E4Geolocation accuracyHow far from the flag did the scout have to search to find it?
- E5Slice evaluationHow does agreement change across crops, cultivars and capture conditions?
- E6Business outcomeHow many walks found nothing, and how much was found only later?
Floor — what the scout finds when they walk it
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 · Capture
2 modes
- SK-01
Degraded capture scored normally
Cloud, low sun or blur is not declared.
- SK-02
Capture change read as crop change
A different sensor or sun angle moves the index.
Stage reads
The imagery, the conditions it was taken in and the field
02 · Detection
2 modes
- SK-03
Detection collapses at a stage
A stage or cultivar seen rarely reads as clean.
- SK-04
Low-prevalence problem missed
A rare pest never reaches the report at all.
Stage finds
The anomalies in the imagery and the extent of each one
03 · Classification
1 mode
- SK-05
Look-alike named as the cause
Herbicide injury is returned as a disease.
Stage labels
The symptom class and the look-alikes that share it
04 · Location
1 mode
- SK-06
Geolocation drift
The scout is sent to the wrong part of the field.
Stage places
Where in the field each flag actually sits on the map
05 · Report
1 mode
- SK-07
Report read as a prescription
A sprayer follows the map with no walk.
Stage returns
The walking order the scout actually reads
- Sev-1 · a wrong cause reaches a spray decision
- Sev-2 · a real problem is not reported
- Sev-3 · the walk is wasted, more to review
Affected slices
Agreement with the walk is not spread evenly across captures
A disagreement rate that reads acceptably across a season of flights can sit almost entirely in the imagery taken in poor light and in the cultivars and stages the model saw 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 |
|---|---|---|---|---|
| Low sun, cloud and blur | 6.4% | 3.4× | Review | |
| Cultivars with thin data | 5.1% | 2.7× | Review | |
| Early growth stages | 3.6% | 1.9× | Watch | |
| Repeat fields, clear midday | 1.9% | 1.0× | Normal |
Bar: walk-disagreement rate lift vs. clear-midday baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What the scout found is the only ground truth there is
A flag is not resolved when the model is retrained. It is resolved when someone walked to it and wrote down what it actually was.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Walk agreement, look-alike confusion and geolocation error are tracked per crop and stage.
- 02
Diagnose
The miss is placed at the capture, the detector, the class chosen or the georeferencing.
- 03
Improve
A class list, capture rule or model release goes out on agronomy sign-off, with its version recorded.
- 04
Verify
Captures the scout already walked are scored again, the overturned flags among them.
- 05
Learn
What the scout wrote becomes the label for that capture, and the look-alike joins the confusion set.
Learn → Detect
The return edge. Ground truth arrives at walking pace, so a cohort is only cleared once the season has put enough confirmed cases in it.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, imagery, detection and location, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Imagery source and sensor assessment.Week 1, Week 2
- 03Boundary, crop and growth-stage mapping.Week 1, Week 2
- 04Georeferencing and capture-quality checks.Week 2
- 05Anomaly detection and symptom classification.Week 2, Week 3
- 06Look-alike sets and out-of-range declaration.Week 3
- 07Change against the previous capture.Week 3, Week 4
- 08Walking order and scout handoff.Week 4
- 09Walk-agreement and confusion evaluation.Week 4, Week 5
- 10Geolocation and degraded-capture tests.Week 5
- 11Farm-system and scouting-app 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 crop, one imagery source | ProductionProduction farm-system integration | AdvancedMulti-crop / multi-sensor fleets |
|---|---|---|---|
| Introduced at Pilot | |||
| Anomaly flagging and geolocation | ✓ | ✓ | ✓ |
| Capture-quality marking | ✓ | ✓ | ✓ |
| Look-alikes named on every flag | ✓ | ✓ | ✓ |
| Cause confirmed by a person in the field | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Change comparison between captures | — | ✓ | ✓ |
| Walking order and scout handoff | — | ✓ | ✓ |
| Farm-system and scouting-app integration | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-crop and multi-sensor coverage | — | — | ✓ |
| Multi-site and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on imagery sources and sensors, farm-system integrations, crop and problem range, evaluation depth, 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
01A season of imagery from the sources you actually fly
→Imagery source, sensor and capture-standard assessment
Week 102Your crop, variety and planting records
→Field boundary, crop and growth-stage data mapping
Week 103Field boundaries and the geometry you trust
→Ingestion, georeferencing and capture-quality checks
Week 204The problems you scout for, in the crops you grow
→Anomaly detection, symptom classes and look-alike sets
Week 305Scouting notes where someone walked and wrote down the cause
→Walk-agreement and class-confusion evaluation
Week 406Flags that turned out to be something else, and the near misses
→Regression cases, look-alike sets and failure-mode testing
Week 407Named agronomists and scouts to walk the pilot fields
→Scout handoff, then supervised walks 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 geolocation testing and the first walks made from a live report.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Workflow discovery, imagery sources and the automation boundary
W2Field geometry, georeferencing and capture-quality checks
W3Anomaly detection, symptom classes and look-alike sets
W4Walk-agreement, class-confusion and geolocation testing
W5Farm-system integration, report format and supervised walks
W6Scouts walk from live reports, then Agent Care starts
- Reading the band
- Week 4 measures the agent against fields your own scouts already walked. Nothing enters a scouting round before that comparison exists.
- At the end of W6
- Reports have run alongside your existing scouting and been walked by the agronomists who read them, 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 scouting agent around the fields you already walk.
Show us a season of imagery, the field records behind it and the scouting notes from the walks that followed. We'll fly one field alongside your scouts and score what came back against what they found standing in the crop.
- Typical delivery
- 4–6 weeks
- Deployment
- Your imagery and farm stack
- Cause confirmed
- By a person in the field, always
- After launch
- Agent Care available
Nestack Agents · Crop-scouting computer vision
AGT-AGR-08 · Agent Care available after launch