Agents

Industry Agent Use Cases

Each pattern combines four inputs: current evidence, scoped history, installed industry knowledge, and explicit objectives and constraints. The agent returns reviewable options and a next safe action—not a promise that a prediction is correct.

Implement any use case with the complete TypeScript/JavaScript, Python, PHP, Go, or Rust reference. Replace the scopes, installed industry pack, objectives, constraints, observations, approvals, and retention policy with the reviewed boundaries for that industry.

Pattern overview

AgentUseful memoryConsequence decisionRequired boundary
Software release agentPrior rollouts, incidents, rollbacks, performance thresholdsDelay, pilot, staged rollout, or releaseRequired checks and release-owner approval
Construction field agentSite conditions, prior delays, inspections, supplier performanceResequence, isolate, repair, or rescheduleQualified site authority and safety policy
Agriculture marketplace agentCrop history, buyer outcomes, quality issues, seasonal logisticsBuyer, fulfillment, storage, or timing optionsCurrent market evidence and commercial approval
Healthcare interoperability agentIntegration incidents and validated resolutionsRoute, retry, isolate, or escalate support workNo clinical decisions; authorized human review
Learner-success agentProgress, completed supports, accessibility preferencesSelect the next reversible supportEducator oversight and protected learner data
Hospitality operations agentGuest preferences, service recovery history, occupancy constraintsRecovery, room, staffing, or escalation optionsPolicy limits and human approval for compensation
Customer-support agentPrior tickets, account commitments, successful resolutionsResolve, escalate, credit, or investigateAuthorization, SLA, and compensation limits
Digital-asset risk agentReviewed wallet events and previous investigationsMonitor, restrict, investigate, or escalateSpecialist risk policy; no autonomous financial action

Software release agent

Install ai-engineering-delivery, Team Code Governance, relevant language packs, and GitHub CI/CD Governance. A useful scope stack is:

industry:ai-engineering-delivery
governance:team-code
delivery:github-actions
service:<service>
history:<service>-releases
tests:release
policy:production-change

The agent recalls verified incidents and previous release outcomes, checks current manifests and CI evidence, then compares reversible release paths. Record build digest, cohort size, error rate, latency, rollback time, approval, and observed customer impact. Future runs can increase uncertainty or shorten the first stage when previous estimates were optimistic.

Construction field operations agent

Install construction-field-ops into industry:construction-field-ops. Keep project, site, supplier, inspection, and safety scopes separate.

Example task:

Compare rescheduling tomorrow’s lift, resequencing indoor work, or proceeding within the approved weather and equipment envelope. Identify missing inspections and stop at the site manager’s approval boundary.

Useful history includes approved delay causes, equipment faults, supplier lead times, inspection results, and actual versus planned durations. Observations can record delay minutes, rework, equipment availability, and safety events.

The agent must not infer structural safety, certify equipment, or replace a qualified site authority. Use current weather, engineering, inspection, and permit evidence from systems that own those facts. Any safety-sensitive reflection remains requires_review before memory write-back.

Agriculture marketplace agent

Install agriculture-market-intelligence into industry:agriculture-market-intelligence.

An agent can combine:

  • crop, variety, grade, quantity, harvest-window, and storage memory
  • prior buyer acceptance, rejection, payment, and logistics outcomes
  • installed market terminology and provenance guidance
  • current price, transport, weather, and availability evidence

It can compare sell-now, store, split-order, alternate-buyer, and alternate-route scenarios. Constraints should include quality expiry, storage capacity, buyer terms, logistics limits, and an accountable commercial approver.

Do not treat stale price memory as a current quote. Record source time and jurisdiction, and require live commercial evidence before a transaction.

Healthcare interoperability support agent

Install healthcare-interoperability-support into industry:healthcare-interoperability-support.

Use it for technical support decisions such as retrying a bounded interface operation, isolating malformed messages, switching to a documented fallback, or escalating to an integration owner. Memory can retain safe summaries of error signatures and validated resolutions without retaining unrestricted clinical payloads.

The agent must not diagnose, recommend treatment, or expose protected health information. Keep technical metadata separate from clinical content, use current interface specifications and organization policy, and require authorized review for actions that may affect care delivery or patient data.

Learner-success agent

Install education-learner-success into industry:education-learner-success.

The agent can remember completed material, demonstrated competencies, accessibility preferences, and previously effective supports. It can compare reversible next steps such as another worked example, a lower-stakes practice set, a different modality, or educator escalation.

Observe completion, comprehension checks, help requests, and educator feedback. Avoid inferring protected traits, ability, diagnosis, or permanent potential. A teacher or authorized support professional owns consequential interventions and any lesson promoted into shared learner policy.

Hospitality and service recovery agent

Install hospitality-guest-operations into industry:hospitality-guest-operations.

Scoped memory can retain approved preferences, prior service issues, promised follow-ups, and recovery outcomes. Consequence can compare room move, maintenance response, alternative amenity, bounded compensation, or manager escalation while considering availability and policy limits.

Record whether the issue was resolved, response time, compensation used, repeat contact, and verified guest feedback. Do not expose another guest’s history or turn one complaint into a permanent preference without confirmation.

Customer-support decision agent

Create customer, product, entitlement, incident, and policy scopes rather than one unrestricted support history.

Example scope stack:

customer:account-123
product:billing-api
incident:inc-456
policy:support-entitlements
history:support-outcomes
agent:support

The agent recalls contractual commitments and similar reviewed resolutions, then compares investigate, resolve, escalate, workaround, or bounded credit options. Hard constraints should prevent unauthorized account changes, unsupported promises, and compensation beyond the caller’s authority.

After resolution, record response time, handoffs, repeat contact, confirmed root cause, and whether the promised action completed. Reflection can reveal that a commonly suggested workaround causes repeated contact and should be demoted in future recommendations.

Digital-asset risk agent

Install digital-asset-risk-wallets into industry:digital-asset-risk-wallets.

Use memory for reviewed wallet labels, prior investigations, source provenance, and policy outcomes. Use Consequence to compare monitor, gather evidence, restrict within existing authority, or escalate. Predictions remain uncertain and must not be represented as proof of ownership, fraud, legality, or investment outcome.

Keep transaction and identity evidence source-linked. Require specialist review, legal and jurisdictional policy, recovery planning, and explicit authorization before any external financial or access-control action.

Design every use case with the same questions

  1. Which Nucleus owns the data and decision?
  2. Which exact scopes may this agent read and write?
  3. What current source is authoritative?
  4. Which installed skill or knowledge base applies?
  5. What options, objectives, constraints, and metrics are bounded?
  6. What would make the agent stop instead of recommend?
  7. Who owns approval and external execution?
  8. Which observed outcome can calibrate the next decision?
  9. Which lesson requires review before becoming memory?
  10. How is the full lifecycle tombstoned when it expires or is deleted?