Each reference agent ships as an OCI container. Inspect the code, modify it, deploy it, or use it as the starting point for your own. All six run under the same human-in-the-loop and autonomy policies.
Drafts replies to inbound tickets and emails, grounded in order history, product policy, and loyalty tier. Reviewer sends, edits, or redrafts. Routine queries flow autonomously once confidence crosses the tenant threshold.
require_review by default. allow_autonomous for low-risk categories at confidence > 0.9.
Nightly three-way match across orders, payment service provider settlements, and bank statements. Classifies discrepancies into nine categories and routes exceptions to the finance team with full lineage.
Autonomous match for 1:1 cases. require_review for category-5+ discrepancies.
Classifies incoming invoices, drafts the amortisation schedule, and routes tier-based approval. Posts to the GL once approved. Learns from reviewer edits to improve classification accuracy over time.
Scope by amount: < £1k autonomous, £1k–£10k require_review, > £10k multi-tier approval.
Drafts SKU copy, structured attributes, and localised variants grounded in brand-voice knowledge. Writes back into the PIM once reviewer approves. Supports batch workflows for new catalogue drops.
require_review for new SKUs. Autonomous for attribute enrichment on existing SKUs.
Daily SKU-level forecast surfaced into the buyer's own planning tools — not a new one. Flags fast movers at risk of stock-out and long-tail at risk of markdown. Learns from reviewer overrides.
Fully review-only — the buyer always decides. Confidence scoring surfaces where the model is unsure.
In-channel answers to policy questions grounded in current handbooks. Audited: who asked, what, when. Routes nuanced cases to a human HR partner with the full context pre-attached.
Autonomous for factual lookups. require_review for anything touching compensation, performance, or grievance.
The agent contract is deliberately thin: an OCI container that speaks Model Context Protocol, with a manifest declaring what capabilities it needs. Build in Python, .NET, Rust, Go, TypeScript — whatever your team knows best.
id: my-returns-agent name: Returns Classifier runtime: kind: container image: ghcr.io/acme/returns-agent:v1.0.0 capabilities: required: - llm.complete - context.get - connectors.ecommerce - connectors.ticketing optional: - vector.search autonomy: default: require_review allow_autonomous: when: confidence > 0.92 unless: refund_amount > 50 triggers: - type: webhook path: /returns/new - type: cron expression: "0 */2 * * *"
45-minute walkthrough with a working customer-service agent — real data, real tickets, real review decisions in the inbox.
