AI agent workflows for Mining, metals & natural resources

Build spec

1 agent — 1 telemetry-evaluation agent; the sensor read, breach check, stop-work approval and CMMS write are workflow nodes.

System prompt (paste-ready)

You are a condition-monitoring assistant for mining equipment. Given an asset's readings {asset_id, vibration, temperature, pressure, run_hours} and its documented alarm/warning limits, classify severity as OK | WATCH | ALARM and return strict JSON {asset_id, severity, breached_signals[{signal, value, limit}], recommended_action, requires_stop: boolean}. Use only the provided readings and documented limits — never invent a threshold or a reading. Set requires_stop=true only when a documented limit for an unsafe condition is breached; you do NOT stop the machine yourself, you flag it for a human to decide.

Guardrails

• The agent never stops or controls an asset — a stop/shutdown recommendation is Approval-gated to the maintenance lead before anything is taken offline
• Severity is derived only from documented limits and the actual readings — no invented thresholds, no estimated values
• Work-order creation is idempotent (keyed by asset + signal + shift) so the same drifting reading can't spawn duplicate tickets in one run

Cost strategy

The evaluation agent runs on the routine telemetry cadence, so it sits on the economy tier — threshold-checking is exactly the cheap, high-frequency case and there is no reasoning premium to pay for it. A per-run tool/token cap bounds cost per telemetry batch, and the Condition gate means most within-tolerance readings stop before any expensive work happens.

Output & delivery

Live readings are read via Device Control → the economy agent classifies severity against documented limits → the Condition gate drops in-tolerance noise → stop-work recommendations wait at the Approval node → a Tool node opens the CMMS work order and the crew is alerted on WhatsApp/SMS; every alert + decision is logged against the asset.

Build spec

1 agent — 1 report-structuring agent; the taxonomy grounding, severity fast-track, safety sign-off and EHS write are workflow nodes.

System prompt (paste-ready)

You are a safety-incident intake assistant on a mine site. Turn a worker's free-text or transcribed spoken report into strict JSON {reported_at, location, asset_or_area, hazard_type, description, people_involved, injury: boolean, provisional_severity} in the site's own terms. Preserve the worker's account faithfully — never downplay, embellish or resolve it, and never assign blame. Classify hazard_type and provisional_severity ONLY against the site safety taxonomy in Knowledge. You do NOT determine the final severity or close the incident — a qualified safety officer reviews and signs off; your output is a draft for that review.

Guardrails

• No incident is closed or filed-as-reviewed without the safety officer Approval — the agent only drafts a structured record; a qualified human classifies and signs off
• Hazard type and provisional severity come only from the site taxonomy in Knowledge — never an invented category, never a severity the rubric doesn't support
• The worker's account is preserved verbatim in the record — the agent never edits away, minimizes or reinterprets what was reported

Cost strategy

Structuring one report and classifying it against a fixed taxonomy is a light, bounded task, so the intake agent runs on the economy/standard tier — there's no need for a premium reasoning model to transcribe and categorize a single incident. Voice reports add a transcription/synthesis step (ElevenLabs/Sarvam) only on the calls that actually come in, so cost tracks real reporting volume.

Output & delivery

The worker's WhatsApp/voice report → the agent drafts a structured record classified against the site taxonomy → the Condition node fast-tracks injuries/high-severity → the safety officer reviews and signs off at the Approval node → a Tool node writes it to the EHS system and the reporter gets a confirmation reference; the draft, the officer's edits and the final record are all logged.