Development practice

Agent operations: from business goals to controlled execution

Operators set goals and rules, Agent prepares analysis and changes, and enterprise systems perform the business operations. A campaign illustrates how workbenches, Skills, data and authorization cooperate.

Design example: a multi-store campaign illustrates responsibilities across operators, Agents and enterprise systems.

Architecture: responsibilities

01

Operators · decisions

Plan, handle exceptions and review outcomes; approve offers, budgets and publication.

Goals, materials and human review ↓

02

Workbench + Agent + Skill

Organize materials and outputs by campaign; Skills define query, analysis and change steps.

Skill-guided authorization, execution and data access ↓

03

Control and authorization

Configuration versions, approvals, permission checks and temporary credentials.

04

Enterprise execution systems

Store, ERP, orders and inventory; access through a proxy or constrained short-lived credentials.

05

Enterprise data layer

Orders, products, channels and campaigns with consistent metrics, timestamps and provenance.

Combine results and evidence for operator review ↓

06

Reusable outputs

Reports, change sets, exceptions and reviews linked to the campaign.

Execution flow

  1. 01

    Define scope

    Select store, dates, metrics and permitted actions; upload the plan and reference materials.

  2. 02

    Read and prepare

    Agent reads authorized data and prepares differences, evidence and a change draft. Search hits are not confirmed facts.

  3. 03

    Approve, then execute

    Bind approval to version, target and action. Check version and permission before execution, then read back results.

  4. 04

    Monitor and review

    Scheduled tasks summarize data and exceptions for review. Retain final configuration, results and explanations after the campaign.

Key design decisions

Separate control from execution

Authorization decides who may act on which store. If the target supports scoped temporary credentials, exchange and connect directly; otherwise use a connector proxy. Both require server-side permission checks; Skill instructions are not authorization.

The data layer owns metric definitions

Connect an existing warehouse or build and host ingestion, cleaning and metric services. Keep periods, currencies, timezones and sources. Queries or programs calculate metrics; Agent explains and asks follow-up questions. Integration and hosting scope depends on the enterprise.

Changes must be reviewable

Use a changeSet containing target, expected revision, before/after values and approval reference. Reuse operation IDs for retries and query uncertain outcomes before repeating writes. Business APIs define compensation; not every action can be rolled back.

Example: changing a gift stock limit

A retailer wants to raise a store’s gift limit from 100 to 120. Agent checks stock and budget and drafts a change. After approval, execution checks that configuration revision is still 12, applies the change and reads it back. A changed revision triggers reconciliation. Proposal, approval and activation remain separately verifiable.

{
  "target": { "storeId": "store-demo", "campaignId": "campaign-demo" },
  "expectedRevision": 12,
  "changes": [{ "field": "giftLimit", "before": 100, "after": 120 }],
  "approvalRef": "approval-demo",
  "operationId": "change-demo-001"
}
Illustrative data contract, not a callable API.

Scope and limits

Use this design for recurring, time-consuming work with reviewable outputs. Business systems and approvals govern activation, spending and publication. A completed Agent task does not prove a campaign took effect.

Continue reading