Agentforce
Gemini is GA in the Agentforce reasoning engine: how to switch, what to test
Gemini has been selectable in Agentforce since May and became generally available in the reasoning engine at Dreamforce. Switching is an org-level setting. Deciding whether to switch takes an evaluation set, a look at your grounding, and a view on where your data and your other Google work already live. This is the practical version: what changed, what to run before you flip it, and when Gemini is the right pick.
Gemini reached general availability inside the Agentforce reasoning engine on September 15, per the Salesforce and Google Cloud announcement. If that sounds familiar, it should. Google’s April press release said Gemini-powered reasoning would be available in May, and the same release counted more than 1,400 customers already using Gemini within Agentforce.
So GA is a status change more than a feature launch. What it changes is the support posture: a generally available model is one you can put a production agent on without a beta caveat in the risk register. Whether you should is the question this post is for.
What GA covers, and what it does not
Three things are now in the GA column. Gemini as a model choice in the Agentforce reasoning engine, the Atlas planning loop that classifies a request, picks a topic and calls actions. Gemini models in Prompt Builder, for prompt templates that generate text rather than plan actions. And, on the Google side, Salesforce reachable from Gemini Enterprise through MCP, with the Agentforce Sales agent listed there since April’s open beta.
Two things are not. Hyperforce on Google Cloud, which is where Salesforce itself would run, reaches general availability in North America in November, with selected US migrations in the fourth quarter and other regions in 2027. And the Data 360 to BigQuery zero-copy expansion, with Iceberg REST Catalog and Private Connect, is dated for fall. I covered the whole partnership in the Google Cloud post; this one stays on the model.
Model selection in Agentforce is at the org level. Read that sentence twice, because it means switching to Gemini is not an experiment on one agent unless your org has a way to scope it, and Salesforce’s published material describes org-level selection across Gemini, Claude and the Salesforce default, with Koa joining from winter as a Salesforce-hosted option.
The model selection post covers how the menu works and what per-subagent overrides exist where they do.
What changes when you switch
The model is the reasoning step. Everything else stays: your topics, your actions, your grounding, your guardrails, your credit metering. A standard action still costs the same number of credits whichever model chose it. What changes is how the model reads your topic descriptions, how it formats tool calls, how it handles a long grounding payload, and how fast it answers.
Those four are where switches go wrong in practice. A topic description tuned by trial and error against one model’s reading of it can classify differently under another. An action whose description leans on phrasing one model happened to interpret well can be called with different inputs. Grounded answers can change in length and confidence.
Latency to first token differs between models and hosting paths, which a customer on chat notices before anyone reads a dashboard.
None of that is an argument against Gemini. It is an argument against switching without a test.
The test to run before you flip it
Build an evaluation set from your own traffic, run it on the current model, switch, run it again, and compare. The set is the same one I described for comparing Koa against your current model, and it is worth building once for every model decision after this one.
Fifty tasks from real sessions, spread across your top intents by volume, your highest-escalation intents and your edge cases. Each with the expected topic, the expected action and inputs, and the expected end state of the record. Ten confidentiality cases where the right answer is refusal. A split between single-turn and multi-turn, read separately.
Run it through the Testing Center, with Custom Scorers encoding action correctness and route correctness as pass-or-fail, and take latency and credits per resolution from the session traces. If the set already lives in your CI pipeline, the switch is a branch.
Then read the four numbers. Gemini earns the switch if action correctness holds or improves, multi-turn does not degrade, ninety-fifth percentile latency stays inside what your channel tolerates, and credits per resolution do not rise. One failure is a reason to stay, whatever the partnership announcement says.
When Gemini is the right pick
Three situations make Gemini the natural choice rather than one option among several.
Your estate is Google-shaped. If your warehouse is BigQuery, your workspace is Google, and your people are already in Gemini Enterprise, the same model on both sides of the MCP bridge means fewer surprises when a workflow crosses from Salesforce into Google and back. The fall zero-copy expansion and November’s Hyperforce move both push the same direction, and co-locating compute and data is the quiet win in that announcement.
Your agent needs to look at things. The April release described Gemini in the reasoning engine as handling text, image and video. If your service agent has to read a photo of a damaged part or a screenshot of an error, that capability is the decision, and it is worth testing with real images from your cases rather than samples.
You want a second GA model to hold against the first. Model risk is a real line in an AI governance policy, and having two generally available reasoning models evaluated against the same set means a price change, an outage or a behaviour regression on one has a tested fallback.
When it is not
Two situations argue the other way.
Your agents are tuned for Claude. Since the Claudeforce partnership, Claude is the default in Coworker, Vibes and Slack, and if your Agentforce agents were built and tested against it, a switch is a retest of everything for a benefit you have not measured. Run the set, and if the numbers tie, stay.
You are waiting for Koa. If your agent’s job is mostly CRM actions, updating opportunities, routing cases and scheduling follow-ups, Salesforce’s own model is due in winter in US regions and claims an edge on exactly that. Switching to Gemini now and Koa in three months is two retests. Switching once, after both are on the same evaluation set, is one.
How to make the switch safely
Do it in a preview sandbox first, on a full copy of the agent’s metadata, with the evaluation set already run on the current model so the baseline exists. Change the org-level model, run the set, and read the four numbers before anyone touches production.
If production must move, move it at a quiet hour with the previous model one setting away, and watch the observability dashboards for escalation rate and latency in the first day. Keep the session traces for at least the first month, because if something drifts, Agent Optimizer will need them in October, and the traces from a model switch are the most useful ones it can read.
And write down why you switched, with the numbers. The next model announcement is a few months away, and the decision will be easier if the last one is documented with the same four numbers, so the comparison is like for like.
Understanding the basics
Is Gemini generally available in Agentforce?
Yes, in the Agentforce reasoning engine, as of September 15, 2026, per the Salesforce and Google Cloud announcement at Dreamforce. Gemini models are also available in Prompt Builder. Gemini had been selectable since May 2026, so GA is a change of support status rather than a new capability.
How do I switch an Agentforce agent to Gemini?
Model selection is at the org level, so the switch applies across the org’s agents rather than one agent in isolation. Test in a preview sandbox first with an evaluation set built from your own traffic, then change the setting in production at a quiet hour with the previous model ready to restore.
Does switching to Gemini change what Agentforce costs?
Not per action. Agentforce meters in Flex Credits per action regardless of model. What can change is how many actions a model takes to resolve a task, so credits per resolution is one of the numbers to measure before and after.
Should I use Gemini, Claude or Koa in Agentforce?
Whichever scores best on your own evaluation set for the agent in question. Gemini suits Google-shaped estates and agents that need to read images. Claude is the default across several Salesforce surfaces since Claudeforce. Koa is Salesforce’s CRM-tuned model, in pilot now and due in winter in US regions.
The setting is easy, the decision is not
Switching a model in Agentforce is one field. The reason it deserves a post is that the field changes how every topic is classified and every action is called, and the only way to know what that does to your agent is to measure it on your agent.
Gemini going GA is good news if you have that measurement. Build the set, run it, and let the numbers make the switch.
If you want the evaluation set built and the model comparison run before you change the org setting, talk to us, or start with the Agentforce readiness assessment.
