Dev.to AI 🤖 Ai 👁 0 📖 5 min read

Part 33: Prompt and model versioning across five platforms

Part 33 · One Agent, Five Platforms A prompt change is a release. So is a model change, and so is a change to the tool list. Series 1 Part 8 made that a rule: each of those changes is a new version, with the same gates

Part 33: Prompt and model versioning across five platforms

Part 33 · One Agent, Five Platforms

A prompt change is a release. So is a model change, and so is a change to the tool list. Series 1 Part 8 made that a rule: each of those changes is a new version, with the same gates as a code change.

This part asks how each of the five platforms lets you version, compare and roll back an agent, and how each one lets you pin a model so it doesn't change underneath you. It's based on vendor documentation I read on 21 November 2026, plus the builds in Parts 11 to 25. Like the builds, it's documentation-only.

Versioning matrix: prompt versions, model pinning, rollback and compare on five platforms

1. What counts as a version

For the invoice agent, a version is three things that change together:

  • The prompt: the instructions and any examples.
  • The model: the provider, the model name and its version.
  • The tool list: which of the nine tools from Part 7 are switched on, and at what stage (Part 9).

If any one changes, the agent's behaviour can change, so the version number changes. A platform that versions only the prompt, and lets the model float, versions a third of the agent.

2. How each platform stores and versions agent config

  • Azure: in Foundry, "each version of an agent is immutable after you save it", according to the development lifecycle page. Code refers to an agent as name:version. Hosted agents get numbered versions too, and the hosted agent page warns that the default endpoint "follows the latest version", so you pin production first.
  • AWS: every AgentCore Runtime update creates a new immutable version. A DEFAULT endpoint follows the latest, and named endpoints stay where you put them (Part 25). Bedrock Prompt Management adds numbered prompt versions, described as a "snapshot of your prompt".
  • Boomi: "A package is a versioned snapshot of your agent configuration," and "you cannot change a package after you create it", according to the packaging page. Versions are numbered automatically from 1.0.
  • MuleSoft: agent networks are published to Exchange, which "handles the semantic versioning and automatically increments the version when there are changes", according to the publishing page. The broker's instructions live in an Agent Script file in the project, so they're versioned with it, as in Git.
  • SAP: the generative AI hub has a prompt registry, with templates referenced "by ID or scenario, template name, and version", according to the SDK release notes. I found nothing on versioning Joule Studio agents themselves, as in Part 18.

3. Pinning model versions

This is where an agent can change without anyone releasing anything.

  • Azure: the model versions page gives three upgrade policies per deployment: opt out of automatic upgrades, upgrade when a new default version is available, or upgrade when the current version expires. Opting out pins the model, but "when the model is retired, those deployments stop working". Azure gives at least two weeks' notice before a new default for Azure OpenAI models.
  • SAP: the generative AI hub course says a model version of latest "will automatically use the most recent version" of that model, and that leaving the version out does the same in orchestration. You can set an exact version instead, and then you "must actively monitor model retirement dates".
  • AWS: Bedrock models are Active, Legacy or End-of-Life, according to the model lifecycle page. Most have a six-month Legacy period before End-of-Life, and migration "will not happen automatically".
  • MuleSoft: the agent network YAML reference lists LLMs by platform (Gemini, OpenAI or Azure OpenAI) and model name. I found nothing on pinning a dated model version, so pinning depends on what the provider's model name allows.
  • Boomi: agents use a Boomi-managed model or your own. I found no model version setting on the pages I read.

The rule: pin an exact model version in production, and put the retirement date in the registry of record from Part 30. An auto-upgrade policy is a release you didn't schedule.

4. Comparing versions on the golden set

A new version should beat, or at least match, the one it replaces on the golden set from Part 10. How much help you get:

  • Azure: the development lifecycle page lets you compare two versions' setup, YAML and chat output on the same inputs, and run evaluations (in preview) "across versions" before publishing.
  • AWS: AgentCore Evaluations scores traces against expected tool calls, which is the closest match to the golden set (Part 25).
  • Boomi: a Test Agent window with traces (Part 15), but no version comparison. You can export each package's configuration as YAML and compare the files yourself.
  • MuleSoft and SAP: no comparison feature in what I read. The golden-set run is your own pipeline, as in Series 1.

5. Rolling back

  • Boomi: documented: "To roll back, select a previous package version" and deploy it to the environment running the newer one.
  • Azure: for hosted agents, pin production to one version, test the candidate, switch, and "restore the previous concrete selector" to roll back. The page adds that sessions stay on the version they started with, and that deleting a version "isn't a rollback operation".
  • AWS: point the named endpoint back to the previous version (Part 25).
  • MuleSoft: no rollback procedure in the docs, as in Part 11. Series 1 redeployed the previous release tag, and kept that version in Exchange until the new one had been stable for a while.
  • SAP: not covered in what I read.

Rolling back the agent doesn't roll back the model. If the deployment auto-upgraded, the old version may already be gone.

6. When a provider updates a model behind the same name

Series 1 Part 9 described the symptom: judge scores falling slowly over a few weeks, with no change on our side. The cause can be a provider updating a model behind the same name. The response was three steps: rerun the golden set, pin a specific model version, and treat any model change as a release.

That still holds on all five platforms. What differs is how easy each makes it to notice. A deployment set to auto-upgrade, or a model version of latest, turns a provider's release into yours.

What I'd take from this

Version the prompt, model and tool list as one unit. Pin the model, keep the previous version deployable, and run the golden set before every switch. AWS, Azure and Boomi give you immutable versions to roll back to; on MuleSoft and SAP, source control and your pipeline do that job.

How do you version prompts today?

This series describes a reference build on a fictional company. This part is based only on vendor documentation read on 21 November 2026 and the documentation-only builds in Parts 10 to 25; no hands-on build has been run yet. Views are my own.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.