Mining AI agent · Autonomous fleet
Autonomous Fleet & Remote-Operations AI Agent
Assemble the context behind every autonomous stoppage, order the intervention queue, prepare pre-start and area-access packs and track zone permissions — while the controller and the supervisor keep every decision that moves a machine.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Machine movement
- Controller
- After launch
- Agent Care
What this agent does
Assembles the picture, never moves the machine
In
- 01
Take the stoppage, the exception and the vehicle state as the control room receives them.
- 02
Read the zone map, the permission state and what changed on the network since handover.
Reason
- 03
Assemble what sat behind the stop — which vehicle, where, what it was doing, what it was permitted.
- 04
Order the open interventions by what holds a person on the ground and what only the controller can clear.
- 05
Prepare the pre-start and area-access packs — the checks, machine states and boundaries a person confirms.
Decide
- 06
Flag a boundary released in the system but not marked on the ground, and permissions that have lapsed.
- 07
Route what it cannot resolve to the controller, the shift supervisor or the statutory position for that area.
Out
- 08
Present the context, the ordered queue and the packs as drafts, with the source records beside them.
- 09
Retain every record read, the zone and network version used, the rationale and every controller amendment.
Product statement
The agent assembles, orders and prepares. Moving a machine, granting a permission, clearing a protective system and declaring an area safe to resume stay with the named person.
Example workflow
One stoppage, end to end
- 1Stoppage receivedAn autonomous vehicle stops, an exception is raised, or an area-access request arrives
- 2State gatheredVehicle, position, task, permission state, zone boundary, network version and what the operator saw
- 3Context assembledWhat sat behind the stop — the sequence, the machines around it and the state each of them was in
- 4Queue and packs preparedOpen interventions ordered, pre-start and area-access packs prepared against the site's own checks
No human action required
Stages 1 to 4 run without a person in the loop — gathering, assembling and ordering finish before anyone is asked to read anything. No machine moves in that stretch.
Reaches the controller as complete context.
Goes to the supervisor before it goes further.
Controller or shift supervisor
Reads the assembled context, the ordered queue, the sources and what is unresolved, then decides what happens to the machine and the area.
Accept context · Amend · Escalate
Context accepted — handed back
- 6Handed to the control roomPresented to the controller as a draft; nothing is dispatched, permitted, cleared or resumed by the agent
- 7Outcome evaluatedWhat the controller corrected, what the ground walk found, and what the interaction reports said by cohort
Amendments
What the controller corrects 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
- Commanding, dispatching or re-tasking an autonomous vehicle.
- Starting, stopping or moving any machine.
- Granting, changing or clearing an area permission.
- Setting, moving or lifting an exclusion zone.
Automation boundary
Agent acts unaided
- ✓Assemble what sat behind a stoppage — the vehicle, the area, the task and the permission state.
- ✓Order the open interventions and what each one waits on.
- ✓Prepare the pre-start and area-access packs.
- ✓Track which people and light vehicles hold a current permission for which zone.
Write actions run only inside the approval boundaries agreed during implementation. A machine movement is never one.
- Overriding or bypassing proximity detection or collision avoidance.
- Clearing an interlock or a protective stop.
- Authorising a person or vehicle into an autonomous zone.
- Declaring an area or a machine safe to resume.
Example output
One stoppage, annotated
Everything the agent assembles is attached to the vehicle, the area and the records it came from.
Stoppage context · one vehicle, one area
Illustrative example- Stoppage as raised
- Haul truck held in exception
- Where
- Ramp intersection, mixed fleet
- Network rev
- Rev 12
- Assembled
- Context and queue order
- Confidence
- 87%
- Resume
- Not decided by the agent
As received
The stop, the place and the network version the control room held at that moment — taken as they arrived.
Evidence used
- Fleet-system event log
- Zone and permission state
- Operator's screen record
Why it is ordered first
The turning loop went live on the last shift and the ground was never marked.
What the score decides
Confidence decides how hard the controller reads this, not whether the truck may move.
Action
Accept contextAmendEscalateValue
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 stoppage and exception
From the fleet system and the control room
Rebuild what the fleet was doing
Use the vehicle states, the zone boundaries, the permissions held and the network version current at the moment it stopped.
Hand over a picture, not a feed
The vehicle, the area, the task and the permission state arrive together, so the controller's reading goes on deciding rather than on collecting.
Order the queue around people
An intervention with a person or a light vehicle on the ground is put ahead of one that is only holding a machine.
Build an evidence trail
Retain the states read, the zone and network version, the order proposed, the confidence and every controller amendment — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Fleet management & AHS
- MineStar Command
- Wenco
- Modular
- AHS event and state feeds
Zones, permissions & network
- Zone maps
- permission registers
- Network changes
- road design
Detection & awareness
- Proximity detection
- site awareness
- Vehicle tracking
- geofences
Agent
Autonomous fleet & remote operations
- Assembles the context
- Orders interventions
- Routes to the controller
Control room & comms
- Control-room consoles
- handover logs
- Network coverage
- comms alarms
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. Context clears every layer before a controller reads it, and anything that moves a machine sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether the context may stand
- L3
- Decides what the agent has no path to
- L4 – L5
- Keep the decision with a person and the record intact
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the whole picture — not only the summary a controller opens.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the context the controller opens
- E1Final-output evaluationWas the context right about the vehicle, the area and the state?
- E2Step-level evaluationDid it read the current zone, permission and network version?
- E3Tool evaluationDid it resolve the correct vehicle, area and event record?
- E4Queue-order evaluationWere interventions with a person on the ground ordered first?
- E5Slice evaluationHow does context quality change across specific stoppage cohorts?
- E6Business outcomeWhat did the controller correct, and what did the ground walk find?
Floor — what actually happened on the ground
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 · Retrieval
2 modes
- AR-01
Boundary live but not marked
A road or loop released in the system was never marked on the ground.
- AR-02
Stale permission or sign-on
A lapsed zone permission is carried into the context as current.
Stage gathers
Vehicle states, zones, permissions and versions
02 · Assembly
1 mode
- AR-03
Exception mode read as stopped
A vehicle holding in exception is described as parked and idle.
Stage proposes
The sequence behind the stop and the queue order
03 · Permission check
2 modes
- AR-04
Manned unit in the path missed
A water cart or dozer working the same ground is not surfaced.
- AR-05
Detection read as complete
Sensor coverage is presented as though nothing could be unseen.
Stage checks
Boundaries, permissions and who is on the ground
04 · Handoff / write
1 mode
- AR-06
Alarm volume floods the queue
So much is raised that the controller stops reading the order.
Stage presents
The context the controller reads and its queue
05 · Change / Version
1 mode
- AR-07
Network change untracked
A zone or road change reaches context with no named approver.
Stage tracks
Model, prompt, zone and network version changes
- Sev-1 · a person could be near a moving machine
- Sev-2 · wrong context reaches the controller
- Sev-3 · context degrades, more is corrected
Affected slices
The risk sits where machines and people share ground
Autonomy is not the variable here; proximity is. A light vehicle on the fleet's ground, a network re-cut this week and a slow link to the control room are what break the context. 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 |
|---|---|---|---|---|
| Mixed-fleet areas, night shift | 5.7% | 3.8× | Review | |
| Newly released network changes | 3.9% | 2.6× | Review | |
| High-latency remote links | 2.7% | 1.8× | Watch | |
| Routine queue and dump stops | 1.4% | 0.9× | Normal |
Bar: corrected-or-escalated context rate, lift vs. routine-stop baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
A near miss changes what the agent must look at
Correcting one record leaves the reason it read wrong in place. What changes is the state the agent reads, the check that passed it, or the ordering rule.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Controller corrections or ground-walk findings move in a cohort.
- 02
Diagnose
What the agent read is put back beside what was actually true at the time.
- 03
Improve
The state test or the ordering rule is changed under change control, by name.
- 04
Verify
Replayed against stoppages the site has already investigated to a cause.
- 05
Learn
The missed state is written into what every later context must resolve.
Learn → Detect
The return edge. What the agent learns changes what it reads, never what a machine is allowed to do.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, fleet and zone sources, context and ordering logic, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Control-room process assessment.Week 1, Week 2
- 03Zone, permission and network-change sources.Week 1, Week 2
- 04Fleet-system event and state data access.Week 2
- 05Stoppage context assembly.Week 2, Week 3
- 06Intervention queue ordering rules.Week 3
- 07Pre-start and area-access pack preparation.Week 3, Week 4
- 08Permission and boundary consistency checks.Week 4
- 09Evaluation suite, cohort tests and regression cases.Week 4, Week 5
- 10Controller review and amendment capture.Week 5
- 11Control-room presentation and write-back.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 area, one fleet | ProductionProduction control-room integration | AdvancedMulti-site / multi-fleet |
|---|---|---|---|
| Introduced at Pilot | |||
| Stoppage context assembled from fleet records | ✓ | ✓ | ✓ |
| Queue ordered around people on the ground | ✓ | ✓ | ✓ |
| Pre-start and area-access packs prepared | ✓ | ✓ | ✓ |
| Permission and boundary disagreements raised | ✓ | ✓ | ✓ |
| Network changes surfaced against the area | ✓ | ✓ | ✓ |
| No command path to a machine or a zone | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Controller review workflow and write-back | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multiple sites, fleets and control rooms | — | — | ✓ |
| Group reporting and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on the fleet and control-room systems in scope, the number of areas and sites, zone and permission sources, 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
01How a stoppage reaches your control room today, and who acts on it
→Control-room process assessment
Week 102Your zone map, permission register and network-change process
→Zone, permission and network-change sources
Week 103Access to fleet-system event and state data
→Fleet-system event and state data access
Week 204Your pre-start checks and how an area is released today
→Pre-start and area-access pack preparation
Week 305The order your controllers actually work interventions in
→Intervention queue ordering rules
Week 306Stoppages and near misses you have already investigated
→Evaluation suite, cohort tests and regression cases
Week 407Named controllers and a supervisor to review context
→Controller review workflow, then pilot shifts and production 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 evaluation work and the first shifts a controller reads context on.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Control-room process, zones and the automation boundary
W2Fleet-system events, permission and network sources wired in
W3Context assembly, queue ordering and the pre-start packs
W4Evaluation suite, cohort tests and permission checks
W5Control-room presentation, supervised shifts and corrections
W6Controllers act from their own board with context beside it, then handover
- Reading the band
- Nothing in week 5 is a live intervention. Controllers read assembled context beside their own board, and the board stays the thing they act on.
- At the end of W6
- Controllers have acted from their own board with the agent's context beside it and what they corrected has been counted, then Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Mining AI agent
Build an autonomous-fleet agent around your control room.
Show us one shift of stoppages, your zone and permission set, and who decides what moves today. If a state could not be resolved from your records the rebuilt context says so, and what the investigation found is what we score it against.
- Typical delivery
- 4–6 weeks
- Deployment
- Your fleet-management stack
- Machine movement
- Controller, always
- After launch
- Agent Care available
Nestack Agents · Autonomous fleet & remote operations
AGT-MN-06 · Agent Care available after launch