Salesforce Flow Benchmark: What a Skill Changes in AI Coding Agents

Last reviewed: October 17, 2026

An AI coding agent builds a Salesforce Flow in a few minutes. Which elements it picks depends above all on its training data. Older patterns such as loops or custom Apex classes are documented there across many years. Newer elements appear far less often, because they have only existed for a few releases. We measured how much a skill changes that choice. Two agents got the same tasks, one with our Flow skill and one without. With the skill, the agent used the Transform element, the HTTP Callout action and Collection Sort. Without it, the agent used none of them, in not a single run.

For a Salesforce team there is more in this than a comparison of two agents. The benchmark shows how strongly an AI coding agent's working environment shapes the result, and why that environment only improves reliably when you measure it. A skill is therefore never finished. After every task, the team notes where the agent worked differently than expected and sharpens the skill accordingly. A benchmark shows whether those changes work. Why agents reach for older Flow patterns in the first place is covered in AI Coding Agents and Current Flow Best Practices.

How we measured

Each task comes from everyday Salesforce project work. Both agents got the same brief and ran the same model. Each agent solved every task three times, each time in a fresh scratch org. Every one of those runs produced a finished flow. The skill under test is the Flow skill in our open-source Salesforce agent skills plugin. The skill gives the agent instructions for current Flow elements, tested example flows and a checker for the metadata it writes.

A newer element only counts in a flow that works. So we deployed every flow to the org and ran automated checks on whether it behaves as asked, such as whether the right records get created with the right amounts. We give every result as a number of runs, such as 3 of 3. The count shows directly whether the agent made the same choice every time.

The tasks are realistic scenarios, and none comes from a customer's org. We don't publish the briefs, seed data, checks or reference solutions, because a public task can end up in a future model's training data. The benchmark would then only measure whether a model already knows that task.

Which Flow elements the agent used with the skill

With the skill, the agent used three elements that no run without it used.

Without the skill, the agent got to the same result the older way. It added up the total inside a loop, sorted the rows in its query and wrote its own Apex class to call the vendor. With the skill, the agent called the vendor's API straight from the flow with the HTTP Callout action. The whole logic therefore stays visible in Flow Builder. Admins without Apex skills can follow and adjust it too.

The skill's instructions and examples made the difference, not the check afterwards. Every run that used the Transform element planned it in its first draft, before the checker ever ran. The API version differed too. With the skill, the agent wrote every flow at the org's current API version. Without it, the agent wrote the older versions 62.0 or 64.0.

None of the three elements is new. Collection Sort shipped in Summer '21, the HTTP Callout action became generally available in Summer '23 and the Transform element in Summer '24. Even so, the agent without the skill took the older route every time. That fits the training data. In documentation, forums and sample projects, loops and Apex classes have been far more common for years than the newer elements. A skill closes that gap without anyone correcting the agent on every task.

Where the skill changed nothing

The agent chose the scheduled path without the skill too. Every run put the contract warning on a scheduled path.

This result helps with maintaining the skill as well. Every instruction the agent reads from a skill takes up room in its context and costs tokens. Instructions for jobs the model already gets right can therefore be cut. Without the comparison against an agent without the skill, this spot would stay invisible.

How reliably the flows worked

Every flow deployed. Except for one run without the skill, every flow passed every check. That one flow, which reports to the vendor, already started when a case was created, not when it was raised to urgent as the task asked. With the skill, the agent first settles which event should start the flow, and only then takes one of the bundled example flows as a template. That way it copies no start condition from an example that doesn't fit its own use case.

What the benchmark measures

The benchmark answers a clearly scoped question. It shows what a skill changes about the choice of Flow elements. The yardstick is Salesforce's own guidance and current Flow architecture. We scoped the three tasks so that every element under study has a clear purpose in them.

The scratch orgs held seed data only and no existing automation. An org with years of history adds another influence. There, the agent reads the existing flows, fields and permissions before every change and takes its cues from them. The benchmark deliberately leaves that influence out, so the skill's effect shows on its own.

What this means for your work with AI coding agents

The benchmark leads to three insights that reach beyond this one skill.

The working environment shapes the result. The same model delivered different flows with the skill than without it. Which solution an AI coding agent delivers therefore depends not only on the model, but also on the instructions, examples and checks it works with. We call that working environment the harness. How a Salesforce team sets up its harness and who owns it is covered in Salesforce AI Engineer.

Every instruction for the agent needs evidence. Only a comparison with the same agent without the skill shows which instructions work, which only cost tokens and which examples send the agent down the wrong path. With three Salesforce releases a year, which element fits a job best also keeps shifting. Skills and project instructions therefore belong under review after every release. That job sits best with whoever owns the harness in the team.

Our Recurring Platform Reviews make visible how this plays out in your org. Among other things, we analyse how consistently your flows use modern elements such as Collection Filter, Collection Sort or the Transform element, and where classic loops still do the work. We also assess how current their API versions are. From one review to the next, that shows whether new automation follows current Salesforce guidance, whether a person or an agent created it.

Your org is the most important context. A skill brings knowledge of current Flow elements. How your org is built, the agent reads from its metadata. Well-described fields, consistent names and current flows are a dependable template for every new change. Older automation, however, serves as a template just as well and then easily gets carried forward. A Platform Assessment shows you where AI coding agents can safely start in your org today, and which areas need a stronger foundation first.

You can try the tested Flow skill right away. It is free in the Salesforce agent skills plugin for Claude Code and Codex.