Skip to content
SteveTechArcSteveTechArc Courses

Orchestrating Agents

Module 8 of 8 in the chapter Integration Orchestration, chapter 15 of the Integration Architect course. Each module stands on its own. Previous: Long Waits and People · Next chapter: Choosing the Right Integration Design

The question this module answers: When an AI agent runs a chain of steps across systems, is it a conductor, and do this chapter’s rules still apply?

A New Kind of Conductor

Section titled “A New Kind of Conductor”

A short look ahead, because this is moving fast and Part 3 of the course takes it in depth. Everything here is dated November 2026; check it again before you build on it.

Give an AI agent a goal and a set of tools, and it decides the steps as it goes: look up the customer, check the balance, place the order, tell the customer. That’s orchestration. The agent is the conductor, choosing and calling the steps.

Salesforce’s own guide to agentic integration describes the same thing: “The orchestrating agent maintains transactional context across the chain, accumulating identifiers, status codes, and data returned by each step.” And it names both directions: other systems’ orchestrators calling Salesforce’s agents (“An external application or AI orchestrator calls an agent programmatically”), and Salesforce’s agents published as services that “external orchestrators or peer agents can delegate tasks to.”

The Same Rules, in Salesforce’s Words

Section titled “The Same Rules, in Salesforce’s Words”

What struck me reading Salesforce’s guidance is how much of it is this chapter, applied to agents.

No rollback across systems (module 3), so design the undo (module 5). The guide: “If full end-to-end success is required, define a compensation strategy for steps that need to be rolled back when a downstream failure occurs.” And: “Design compensation paths as named actions alongside the forward path.” That’s the saga: an undo step for every step that can be taken back, built before you need it.

Every write safe to repeat (chapter 14). The guide gives a reason that’s new with agents: “All write actions must be idempotent; the agent’s reasoning loop may invoke the same action more than once due to Large Language Model (LLM) retries or ambiguous confirmations.” A human conductor sends a request twice by accident now and then. An agent’s reasoning can do it as a matter of course.

The unknown answer (module 1). “If a write action potentially succeeded, but returned an ambiguous response (network timeout, no acknowledgment), the retry must use the same idempotency key or external reference ID as the original call. Never issue an unchecked retry on a non-idempotent write.” Ask before you act again, with the same key.

The rules didn’t change because the conductor got smarter. They got more important, because the conductor is less predictable.

The Agent as the Conductor, or as a Step

Section titled “The Agent as the Conductor, or as a Step”

That leads to the architect’s real choice, and it’s module 6’s question with a new candidate: where does the conductor live, and should it be an agent?

The agent as the conductor, or as a step The agent as the conductor, or as a step: two panels. The agent as the conductor: it decides the steps as it goes; flexible, good when no two requests are the same; but it may choose different steps tomorrow, so every action must be safe to repeat, and every undo named and ready. The agent as a step: a designed conductor, a flow, the middleware or a workflow engine, runs the fixed steps and the undo steps; the agent is called for the part that needs judgment, such as reading a request or drafting a reply; the steps and the undo steps are the same every time. Underneath: let the agent decide what needs judgment; let a designed conductor own what must happen the same way every time. {"author": "Steve Simpson", "blueprint": "1.0.0", "builder": "standard 1.58.0", "built": "2026-11-04", "id": "DGM-INT-OR-14", "rights": "\u00a9 2026 Twinshark LLC. All rights reserved.", "template": "comparison/side-by-side 0.5.0", "theme": "stevetecharc 1.4.0"} The agent as the conductor it decides the steps as it goes Flexible: good when no two requests are the same. It may choose different steps tomorrow. So every action must be safe to repeat. And every undo named and ready. The agent as a step a designed conductor calls it A flow, the middleware or an engine runs the fixed steps. The agent does the part that needs judgment. The steps and undo steps are the same every time. Let the agent decide what needs judgment. Let a designed conductor own what must happen the same way every time. © 2026 Twinshark LLC · stevetecharc.com
Click to enlarge

Judgment to the agent; the fixed path to a designed conductor.

The agent as the conductor fits work where no two requests are alike, and the steps really do depend on what the agent finds. The cost is predictability: it may take a different path tomorrow, so every action it can call has to be safe to repeat, and every undo has to be there for it to use.

The agent as a step fits work where the steps must be the same every time, especially the undo steps. A designed conductor, a flow, the middleware or a workflow engine (module 6), runs the fixed path, and calls the agent for the part that needs judgment: reading a customer’s email, drafting the reply, deciding which of three options fits.

Integration platforms are adding pieces for this. MuleSoft, Salesforce’s own, describes its Agent Broker, generally available by its 2026 mid-year release, as “a context-aware routing service that discovers and engages the best-fit agents and tools for the job.” Other platforms have their own; chapters 22 and 23 look at the landscape, and at how agents call each other.

Beyond the exam: what AI changes. This whole module is that box. The short version: an agent doesn’t remove the need for orchestration design; it makes it more important. Every step it can call needs an undo, every write needs to be safe to repeat, and every wait needs a deadline, exactly as for any other conductor.

What an Architect Decides Here

Section titled “What an Architect Decides Here”
  • For each process: an agent as the conductor, or as a step inside a designed one, and why.
  • For every action an agent can call: its undo, named beside it, and whether it’s safe to repeat.
  • The key the agent must reuse when an answer is unclear.
  • When to check this again: the platforms are changing every release.

Explore

Section titled “Explore”

Follow Me

© 2023–2026 Twinshark LLC. All rights reserved. · Content Use

Privacy notice

SteveTechArc is independent and is not affiliated with, sponsored by, or endorsed by Salesforce, Inc. Salesforce and related marks are trademarks of Salesforce, Inc.

Educational content, shared as is. Not professional, legal or financial advice. Salesforce changes several times a year, so check current Salesforce documentation before you rely on any platform detail.