Skip to main content
Choose and fund

Agents and packaged applications

Most processes run inside packaged software you cannot rewrite. Two tests decide whether to wait for the vendor's agent, buy it, or build a thin one.

By Murali SidLinkedInPublished Updated

Most of the processes an enterprise wants to automate run inside packaged software it cannot rewrite. Billing, finance, asset and field systems are bought, not written. So the question is not only whether to build an agent. It is also whether the vendor will ship one for your edition, and when. Two tests answer it. Is the process inside one package or across several? Is a vendor agent generally available on your edition, or dated within two quarters? Inside one package, with a vendor agent available or dated, wait or buy. Across packages, or on an excluded edition, build a thin agent of your own.

The two tests

Two questions come first. Does the work need an agent at all, or will ordinary code do? Can its result be judged? Choose where to start covers both.

A decision flow. Does it need an agent? If not, use ordinary code. Can it be evaluated? If not, prepare first. Then come the two tests. Is the process inside one package? Is a vendor agent generally available on your edition, or dated within two quarters? If yes, wait or buy. With vendor agents in several packages, use and govern them. Otherwise build thin. Every agent then gets an impact score on the risk profile and a portfolio score.
Two tests: wait, buy or build thinWait for a vendor only when a dated agent exists for your edition. Build thin where the work crosses packages.Open image at full size ↗
  1. Is the process inside one package or across several? Inside one package, the vendor holds the data and the permissions, and is likely to ship an agent. Across several packages, no vendor agent covers the whole job.
  2. Is a vendor agent generally available on your edition, or dated within two quarters? Generally available means any customer can buy it and run it in production. Count only agents marked that way in release notes. Treat an announced agent as unavailable. Check the price after any free period ends.

Wait, buy or build thin

Thin means small and easy to retire. The agent keeps no business rules of its own. When a vendor agent later covers the job, you switch yours off. The interfaces and definitions stay.

Then treat every agent the same, bought or built. Answer the four questions for it. Score its impact, meaning what it can damage, on the risk profile. Check which deterministic zones it touches. A bill change is a money action, whoever built the agent. Then score it for the use-case portfolio.

Where vendor agents are today

As of 14 October 2026, every large package vendor sells agents. Most need the vendor's cloud edition. Oracle's agents are not coming to E-Business Suite, JD Edwards, PeopleSoft or Hyperion. Prices are still moving. Dates slip. An SAP billing agent for utilities moved from Q2 to Q3 2026 between two articles [vendor]. No vendor publishes how many of its agents run in production.

A worked example: a utility's packaged systems

Take a typical electricity or water utility. The utilities page shows its full estate, layer by layer, as do healthcare and energy. The table below picks out its packaged systems.

Seven packaged systems in a utility estate, badged with what vendors ship as of 14 October 2026. Generally available vendor agents exist for BI, portals, integration, grid systems and SAP billing self-service. SAP billing exceptions, SAP's MCP Gateway and Microsoft scheduling are planned but not confirmed. Nothing was found for GIS, asset registers or Oracle billing. A band marks the join across billing, outage, GIS and field. A strip at the base lists what you own: the interface inventory, business definitions, agent registry and gateway policy.
Where agents land in a packaged-application estateVendors cover most single systems. The join across systems is the place to build.Open image at full size ↗

No measured result exists yet for an agent inside either billing system, or for a portal agent an enterprise built. The one measured result on the line of business and OT page is for the asset-data work. No utility found lets an AI agent operate protection or switching.

Take an outage. Billing knows who is affected, the outage system what failed, GIS where, and the field system who is coming. A customer message and a work order need all four. No single vendor's agent sees all four today. That join is the strongest place to build. Gartner expects agent networks across applications only in 2028 (as reported 27 August 2025).

What to own whatever you buy

Four things stay yours whichever agents you run. They also make each vendor agent easier to replace.

  • Interface inventory. List every interface between systems, with its owner and purpose. Retire dead flows. Decide which ones become agent tools.
  • Business definitions. Keep one set of business definitions under every reporting agent, often called a semantic layer.
  • Agent registry. Keep one registry of your own agents and vendor-embedded agents. Agents you build run through your gateway. Vendor-embedded agents run on the vendor's platform, and you govern them through its settings. Several vendors offer to hold this list [vendor]. It must hold other vendors' agents too. The agent lifecycle lists the record's fields.
  • Policy at the gateway. Put access and approval rules at your own tool gateway and policy decision point, not inside each vendor's agent. Then a change of vendor does not change the rules.

Evidence and limits

The evidence does not settle wait versus build. The case against building on middleware, portals and reporting rests on documented problems. Poor data limits any agent, as SAP's chief executive said in May 2026 [vendor]. Old systems lack modern APIs and secure identity (Deloitte, 10 December 2025). Vendors are building into the same places (SAPinsider, 13 August 2026). The case against waiting rests on dated facts about editions, prices and slipping dates. Lock-in is analyst opinion (Futurum, 12 June 2026).

The sources stop short in several places. No named custom agent was found that a vendor feature later replaced. No measured outcome exists for a portal, middleware or reporting agent an enterprise built. No published framework found asks either test. Microsoft's Cloud Adoption Framework comes closest (updated 2 June 2026) [vendor]. The two tests and the two-quarter window are a proposal, not yet tested against results.

Next step

Put a value and a full cost on the agent on measure value and full cost.

Explore this topic

Vendor status as of 14 October 2026.

  • SAP. Over 30 agents and 2,500 Joule skills (14 April 2026) [vendor]. AI tied to cloud editions or a committed migration. Joule Studio free only through 2026. No general availability date for the "Autonomous Suite" (12 May 2026) [vendor].
  • Oracle. 1,000-plus prebuilt agents (Forrester, 8 May 2026) [vendor]. They need the Fusion cloud applications. Renewal pricing unresolved.
  • Salesforce. Agentforce annual recurring revenue of USD 1.2 billion, quarter to 30 April 2026 [vendor]. Pricing moved from USD 2 per conversation to per-action credits.
  • Microsoft. Dynamics 365 agents slated for April to September 2026 (18 March 2026) [vendor].
  • Registries. SAP, ServiceNow, Workday, Microsoft and MuleSoft each offer to list other vendors' agents [vendor]. Workday's registry is not confirmed as generally available.

Gaps. Gartner's notes were seen only as abstracts. No source was found for Oracle's utilities agents, or for a per-agent list of what shipped at Sapphire 2026.

Sources.

Cite this page

Written by Murali Sid, Founder and AI Systems Architect, last updated 15 October 2026. Licensed CC BY-SA 4.0, so quote it freely with attribution.

Murali Sid (2026). Agents and packaged applications. Agentic Architecture Skills. https://www.agenticarchitectureskills.com/use-cases/packaged-applications

Machine-readable version of this page: /use-cases/packaged-applications.md

On this page