Agriculture AI agent · Autonomous equipment
Autonomous-Equipment Supervision Agent (Watch & Alert)
Watch a fleet of automated machines through a field operation — progress, telemetry, boundaries and stops — and page the named supervisor, who keeps control of every machine and answers for the operation.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Machine control
- Supervisor only
- After launch
- Agent Care
What this agent does
Watches the fleet, never drives it
In
- 01
Take the operation as planned — machines assigned, the work order, the field file, boundaries and exclusion zones.
- 02
Read machine and implement telemetry: position, speed, implement state, coverage, fuel and consumable level.
Reason
- 03
Compare the run against the plan — progress, coverage and overlap, boundary adherence and where each machine sits.
- 04
Watch what the machines report: stops, obstacle-detection events, interventions, faults and gaps in telemetry.
- 05
Read interventions in aggregate, so repeated stops in one place are visible as a pattern and not as single events.
Decide
- 06
Separate what needs a person now — a stop, a boundary event, a silent machine — from what waits for the shift report.
- 07
Treat detection as incomplete: dust, low light, children and people lying down are known limits, and alerts say so.
Out
- 08
Page the named supervisor by the agreed path and keep escalating until a person acknowledges the event.
- 09
Record the event, the machine state, the telemetry around it, who was paged and who answered.
Product statement
The agent watches, alerts and records. Control of every machine, and responsibility for the operation, stays with the named supervisor.
Example workflow
One field operation, end to end
- 1Operation startedThe work order, the machines assigned, the field file, boundaries and exclusion zones
- 2Machines readPosition, speed, implement state, coverage, fuel and consumable level, machine by machine
- 3Events watchedStops, obstacle-detection events, boundary events, interventions, faults and telemetry gaps
- 4Run checked against the planProgress, coverage and overlap, boundary adherence, and interventions read in aggregate
No human action required
Stages 1 to 4 run continuously, not when somebody opens a screen. The agent watching is not the supervisor watching, and a quiet lane is not permission to look away. A stop or a boundary event ends the stretch the moment it is seen.
Raised in the shift report, not paged.
Pages the named supervisor immediately.
Named supervisor
Reads the event, the machine state and the telemetry around it, then decides what happens to the machine and the field.
Attend · Call the operator · Stand the fleet down
Attended — handed back
- 6Alert raised and recordedSent by the agreed path and repeated until a person acknowledges; no command reaches a machine
- 7Outcome evaluatedTime to acknowledgement, missed and late escalations, false alerts and load by field condition
Interventions
Every stop and every intervention is counted, and read in aggregate for a developing fault.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Clearing a stop or restarting a machine.
- Sending any command to a machine or implement.
- Moving, widening or switching off a boundary.
- Deciding a detection event was a false alarm.
Automation boundary
Agent acts unaided
- ✓Watch progress, coverage, telemetry and implement state against the plan.
- ✓Raise stops, obstacle-detection events, boundary events and telemetry gaps.
- ✓Escalate by the agreed path and keep escalating until a person acknowledges.
- ✓Record what happened, who was paged, who answered and how long it took.
Write actions run only inside the boundaries agreed in implementation. No control path to a machine exists.
- Judging a field fit for autonomous work.
- Authorising road travel between fields.
- Setting how many machines one person supervises.
- Changing alert rules, thresholds or escalation paths.
Example output
One stop event, annotated
Everything the agent sends is attached to the machine, the moment and the telemetry it came from.
Alert output · single machine event
Illustrative example- Machine
- Autonomous tractor, unit 4
- Event
- Stopped itself, headland pass
- Conditions
- 21:06, low light
- Sent to
- Named supervisor, paged
- Confidence
- 93%
- Machine state
- Stopped by its own system
As received
The event as the machine reported it, with the telemetry either side of it.
Evidence used
- Machine stop code
- Position and boundary
- Last telemetry, 40s
Why a person was paged
It stopped itself and stayed stopped, and nothing in the telemetry says why it did.
What the score decides
Confidence decides how the alert is worded, not whether it is sent at all.
Action
AttendCall the operatorStand the fleet downValue
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 machine in the run
From telemetry, the work order and the field file
Apply the plan that governs the run
Use the work order, the current field file, the boundary and exclusion zones, and what each machine reports — none of which shows what a sensor missed.
Get the event to a person in time
A stop, a boundary event or a machine gone quiet reaches whoever is on the rota, by name, instead of waiting for the next look at a screen.
Show the pattern, not just the event
Interventions that look ordinary one at a time — three stops in the same corner, a creeping fuel burn — are read together and put in front of a person.
Build an evidence trail
Retain the event, the machine state, the telemetry around it, the field file in force, who was paged, who answered and when — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Machines & telemetry
- Autonomous tractors
- implements
- Machine telemetry
- as-applied data
Fleet & field files
- Operations Center
- FieldView
- Trimble Ag
- boundaries and zones
Alerting & on-call
- Phone and SMS
- on-call rota
- Push and radio
- escalation paths
Agent
Autonomous-equipment supervision
- Watches the fleet
- Raises the event
- Records the answer
Conditions & site
- Weather and light
- dust
- Cameras
- gate and road status
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 fleet
Each control wraps the one inside it. An event clears every layer before a person is paged, and none of the six is a substitute for the machine's own safety systems.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Raise every event, and hold no path to a machine
- L3
- Decides that an alert reaches a person
- L4 – L5
- Keep the machine with a person and the record intact
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the whole watch — not only the alert that got sent.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the alert the supervisor reads
- E1Final-output evaluationWas the event real, and did it describe what the machine did?
- E2Step-level evaluationDid it read the current field file, boundary and machine state?
- E3Tool evaluationDid it poll the right machine, feed and telemetry window?
- E4Escalation timingHow long from the machine event to a named person acknowledging it?
- E5Slice evaluationHow do detection and escalation change by field condition?
- E6Business outcomeHow many events were missed, late, or raised for nothing?
Floor — whether a person actually attended
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 · Telemetry / intake
2 modes
- EQ-02
Telemetry gap reads as normal
A silent feed looks like a machine still working.
- EQ-03
Stale boundary file
An exclusion zone changed after the file was pulled.
Stage gathers
Machine state, position, implement and the field file
02 · Watching
1 mode
- EQ-04
Pattern only visible in aggregate
Repeat stops in one corner never add up to a fault.
Stage compares
The run against the plan, the boundary and the coverage
03 · Alerting
2 modes
- EQ-01
Stop surfaced late
A machine sits stopped with nobody paged.
- EQ-05
Alert flood dulls the watch
So much is raised that alerts stop being read.
Stage raises
The event, who it names and how far it escalates
04 · Output
1 mode
- EQ-06
Detection limits not stated
A quiet screen is read as a clear field.
Stage returns
What the supervisor sees when the alert arrives
05 · Change / Version
1 mode
- EQ-07
Silent feed regression
A feed or field-file change moves what is watched.
Stage tracks
Model, prompt, field-file and telemetry-feed changes
- Sev-1 · a hazard or a stop goes unattended
- Sev-2 · the supervisor is misled about the field
- Sev-3 · watching degrades, more falls to people
Affected slices
Watching does not hold up evenly across field conditions
A late-escalation rate that reads acceptably across a season sits mostly in the conditions where sensing and connectivity are worst — dust, darkness and broken field edges. 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 |
|---|---|---|---|---|
| Dust and harvest conditions | 8.4% | 3.8× | Review | |
| Low light and night work | 6.0% | 2.7× | Review | |
| Irregular and small boundaries | 4.2% | 1.9× | Watch | |
| Daylight single-machine work | 1.7% | 0.8× | Normal |
Bar: late-escalation rate lift vs. daylight baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
The field is put right first, then the watch is changed
A machine standing in a field is not a ticket. It is attended and the field made safe first — the loop only decides what the next shift is watched with.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Late escalations, missed events and false alerts cluster by field condition first.
- 02
Diagnose
The feed, the field file, the watch rule and the escalation path are each ruled in or out.
- 03
Improve
Your safety lead approves the changed rule, source or routing, and the shift it starts on is recorded.
- 04
Verify
Recorded runs from the same conditions are replayed, including the events nobody answered.
- 05
Learn
The event is added to the replay set, and the next shift briefing names what changed.
Learn → Detect
The return edge. Changing what is watched or who is paged is a safety change — your safety lead signs it before a shift runs on 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, feeds, watching logic, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Machine, implement and telemetry feed assessment.Week 1, Week 2
- 03Supervision model, rota and escalation paths.Week 1, Week 2
- 04Field files, boundaries and exclusion zones.Week 2
- 05Event and telemetry ingestion.Week 2, Week 3
- 06Progress, coverage and boundary watching.Week 3
- 07Stop, intervention and telemetry-gap detection.Week 3, Week 4
- 08Alerting, acknowledgement and escalation workflow.Week 4
- 09Escalation-timing and missed-event evaluation.Week 4, Week 5
- 10False-alert and supervisor-load testing.Week 5
- 11Fleet-system, alerting and record 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 machine, one field | ProductionProduction fleet integration | AdvancedMulti-fleet / contractor operations |
|---|---|---|---|
| Introduced at Pilot | |||
| Stop and telemetry-gap detection | ✓ | ✓ | ✓ |
| Alerting to a named supervisor | ✓ | ✓ | ✓ |
| Machine control left with the supervisor | ✓ | ✓ | ✓ |
| Acknowledgement and escalation record | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Progress, coverage and overlap watching | — | ✓ | ✓ |
| Boundary and exclusion-zone adherence | — | ✓ | ✓ |
| Intervention-pattern reporting | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-fleet and contractor operations | — | — | ✓ |
| Shift handover and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on machine types and telemetry access, fleet-system integrations, field and boundary sources, alerting paths, the supervision model, 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 machines, implements and what each one reports
→Machine, implement and telemetry feed assessment
Week 102Who supervises, on what rota, and how they are reached
→Supervision model, rota and escalation paths
Week 103Field boundaries and exclusion zones you actually trust
→Field files, boundaries and exclusion-zone checks
Week 204What each machine's own safety system does when it stops
→Event detection mapped to what the machine already does
Week 305Runs that went wrong, including the near misses
→Evaluation suite, regression cases and failure-mode testing
Week 406What your alerting can carry, and what already gets ignored
→False-alert and supervisor-load testing
Week 507A named supervisor and a safety lead to review
→Acknowledgement workflow, then supervised shifts 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 load testing and the first supervised shifts.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Workflow discovery, the supervision model and the automation boundary
W2Telemetry ingestion, field files and boundary geometry
W3Progress, coverage, stop and telemetry-gap watching
W4Escalation-timing, missed-event and false-alert testing
W5Fleet integration, supervisor-load testing and supervised shifts
W6Supervisors work from live alerts, then Agent Care starts
- Reading the band
- Escalation is tested in week 4, before a supervisor works from an alert in week 5. Nothing is relied on in a working field until it has.
- At the end of W6
- Alerts have run alongside your existing radio and phone practice and been answered by the people named on the rota, 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 supervision agent around your fleet and your rota.
Show us a week of machine data, your field files and how a stop reaches a person today. We'll run a shadow week alongside your rota, paging nobody, and compare what we would have raised against what was actually seen.
- Typical delivery
- 4–6 weeks
- Deployment
- Your fleet and telemetry stack
- Machine control
- Named supervisor, always
- After launch
- Agent Care available
Nestack Agents · Autonomous-equipment supervision
AGT-AGR-07 · Agent Care available after launch