Intent Taxonomy

The advisor cannot route to the right page unless intents are enumerated. Every user message is classified into one of seven intents. Each intent has a fixed contract: entry surface → data reads → target page → overlay → allowed artifacts → close-the-loop event.

Intent classification happens in the Interpret stage of the agent loop. If a message does not classify cleanly into one intent, the advisor asks a scoping question rather than guessing.

The seven intents

Intent	Trigger phrase example	Primary artifact	Target surface
Inform	”What’s happening in Kedah?”	Answer or View	Regional Forecast, National Overview
Diagnose	”Why is Blok A2 stressed?”	View	Field page with risk card expanded
Decide	”Should I irrigate Blok A2 today?”	Answer with recommendation	Chat (stays in place)
Act	”Assign scouting to Ali for Blok A2”	Task	Field Scouting draft
Verify	”Did the last irrigation work?”	View	Verification with before/after
Learn	”Show me last season’s cycle for this estate”	View	Cycle Analysis
Configure	”Mute alerts on Blok A2 for the weekend”	Task	Notification Preferences draft

Per-intent contracts

Inform

Purpose: user wants situational awareness. No decision required, no action requested.

Field	Value
Entry surface	Chat launched from anywhere
Reads	Aggregation Model rollups, Regional Forecast, National Overview, current alerts
Target page	Regional Forecast, National Overview, or Estate Group depending on scope
Overlay	Highlighted regions of interest, reasoning ribbon (“why this page”)
Allowed artifacts	Answer, View
Close-the-loop	None (read-only)
Failure mode	If scope is ambiguous (“how are things?”), advisor asks: national, regional, or estate?

Diagnose

Purpose: user wants to know why a specific entity is showing anomalous behavior.

Field	Value
Entry surface	Chat launched from a field, block, or alert
Reads	Field data, Indices, Crop Cycle Models, firing Risk Model rule cards, matching crop literature diagnosis page
Target page	The field’s page, with the firing rule card expanded and diagnosis excerpt inline
Overlay	Stressed sub-zones highlighted, reasoning panel with `drivers` and confidence, suggested-action strip
Allowed artifacts	View, then optionally Task (offer scout task or VRA draft)
Close-the-loop	If user accepts a follow-on Task, standard scout or VRA lifecycle
Failure mode	If no rule card matches, advisor returns an Answer citing raw indices and offers to file a Proposal for a missing pattern

Decide

Purpose: user wants a recommendation. Advisor produces a bounded recommendation with confidence and stops.

Field	Value
Entry surface	Chat, typically after a Diagnose exchange
Reads	Same as Diagnose, plus `mitigation.actions` and `mitigation.window_days_by_severity` from the matching rule card
Target page	Stays in chat; does not navigate
Overlay	Reasoning panel with recommendation, confidence, alternative options
Allowed artifacts	Answer
Close-the-loop	None until the user escalates to Act
Failure mode	Confidence below threshold → advisor recommends Diagnose or scout first, does not fabricate a call

Act

Purpose: user wants a concrete write. Assign, draft, mute, schedule.

Field	Value
Entry surface	Chat, or from a Decide response
Reads	Target module’s schema, entitlements, current state of the target entity
Target page	The target module’s draft view (Field Scouting, VRA Maps, Notification Preferences)
Overlay	Pre-filled form, module guardrail validation banner, explicit confirmation button
Allowed artifacts	Task
Close-the-loop	Task-created event into Activity & Alerts; downstream module lifecycle takes over
Failure mode	Module refusal (e.g., assignee lacks field access) surfaces the refusal reason; advisor offers an alternative that would pass

Verify

Purpose: user wants proof of outcome after an action.

Field	Value
Entry surface	Chat, or from a completed task in the Activity feed
Reads	Verification Model, source events (scout reports, as-applied maps), satellite passes in the tolerance window
Target page	Verification with the relevant bundle open
Overlay	Before/after comparison, cross-check verdict, source-event list
Allowed artifacts	View
Close-the-loop	Approved bundles are exportable; overrides trigger escalation per Verification guardrails
Failure mode	Insufficient imagery → advisor states the reason and offers to schedule the next check; does not override

Learn

Purpose: user wants historical or educational context, not a live decision.

Field	Value
Entry surface	Chat
Reads	Cycle Analysis, Activity Log, historical rollups, crop literature
Target page	Cycle Analysis, Activity Log, or a diagnosis page
Overlay	Highlighted period, narrative summary, links to related literature
Allowed artifacts	Answer, View
Close-the-loop	None (read-only)
Failure mode	Incomplete cycle → advisor states what is missing and offers a partial narrative

Configure

Purpose: user wants to change platform behavior for themselves or the estate.

Field	Value
Entry surface	Chat
Reads	Current preferences, entitlements, safety-floor rules
Target page	Notification Preferences, Estate Group settings
Overlay	Diff preview (before/after), safety-floor banner if applicable, explicit confirmation
Allowed artifacts	Task
Close-the-loop	Preference-changed event into audit log; downstream alert routing updates
Failure mode	Attempt to disable a safety-floor channel → module refusal; advisor explains and suggests the bounded alternative

How intent selects the artifact

Intent	Answer	View	Task	Proposal
Inform	✓	✓
Diagnose		✓	offered as follow-on	offered when no match
Decide	✓
Act			✓
Verify		✓
Learn	✓	✓
Configure			✓

Proposal is not tied to a single intent. It can be issued from any intent when the advisor detects that the knowledge base is missing a pattern (a new stressor, a new mitigation, a threshold that fires too aggressively). Proposals always route to human review; they never auto-apply. See the knowledge tiers.

Ambiguity handling

If the classifier confidence is below threshold or two intents are close, the advisor asks a single scoping question before proceeding. It does not guess and it does not chain multiple clarifiers. Example:

User: “What about Blok A2?” Advisor: “Do you want a diagnosis, a scout assignment, or a historical cycle view?”

This keeps the loop deterministic and testable.

Multi-intent messages

A message like “Blok A2 looks stressed, assign Ali to scout it and mute the alerts until he reports back” contains Diagnose + Act + Configure. The advisor executes them as a sequence of separate loops, each with its own artifact and confirmation, in dependency order:

Diagnose → View on Blok A2 (auto-confirmed, read-only).
Act → Task: scout assignment to Ali (human confirms).
Configure → Task: bounded alert mute (human confirms; safety floor still applies).

Each step’s output is available as context to the next. Nothing is bundled into a single opaque action.

Upcoming pages will cover:

Overlays and navigation — the UI primitives that make View artifacts feel like the agent is driving the page.
Preset playbooks — named end-to-end compositions of intents (Morning Brief, Diagnose a Field, Assign a Scout, Verify a Treatment, Plan a VRA Map, Explain an Alert, Rollup Review, Weekly Cycle Debrief).
Agentic actions and guardrails in detail — the full matrix of what the advisor can do against every module.
Handoff and escalation — chat → scout task → assignee → activity log → verification.

Get Started

Core Concepts

Guides

Integrations

Help

The seven intents

Per-intent contracts

How intent selects the artifact

Ambiguity handling

Multi-intent messages

Next

​The seven intents

​Per-intent contracts

​How intent selects the artifact

​Ambiguity handling

​Multi-intent messages

​Next

The seven intents

Per-intent contracts

How intent selects the artifact

Ambiguity handling

Multi-intent messages

Next