Agentforce
Salesforce and AWS after Dreamforce 2026: Bedrock models, Amazon Quick, and the Connect handoff
The Google Cloud announcement got the keynote minutes, but the AWS one changes more for orgs that already run on AWS. Agentforce can now call Anthropic and NVIDIA models through Bedrock, Salesforce data shows up inside Amazon Quick, AWS's own agents arrive in Slack, and a voice handoff between Amazon Connect and Agentforce Voice is due this fall. This is what shipped, what is dated, and what to do with each.
Most of the Dreamforce coverage went to Google: Hyperforce moving onto Google Cloud, Gemini inside the reasoning engine. The AWS announcement landed the same day and got a fraction of the attention, which is odd, because for an org that already has its warehouse, its contact centre, or its AI spend on AWS, it is the one with more to act on.
I read the AWS and Salesforce announcement as seven separate items wearing one press release. Three are available now, three are dated for the fall, and one is procurement rather than product. Sorting them that way tells you what to do this month and what to put on a roadmap.
What is available now
Three things you can use today, if your contracts allow.
Agentforce can call Anthropic and NVIDIA models through Amazon Bedrock. OpenAI models are listed as coming soon. Salesforce says the Bedrock path keeps zero data retention enforced and carries HIPAA, PCI, SOC 2 and ISO 42001 compliance, with dynamic grounding and toxicity detection still applied.
For an org whose security review has already approved Bedrock, this removes the argument about a second model provider. The model menu inside Agentforce now has an AWS-hosted lane alongside the Salesforce-hosted one, and if you have wanted a bring-your-own model route that your cloud team already governs, this is it.
Amazon Quick, AWS’s assistant for work, can read Salesforce data natively. Pipeline health, account summaries, service cases and operational analytics show up inside Quick, and the connection runs on the Model Context Protocol rather than a bespoke integration.
If you have read about the Headless 360 MCP server, Quick is that server with a new client on the other end. Quick becomes another AIforce surface, in Salesforce’s own framing, and the same permission rules apply: a user sees in Quick what they can see in Salesforce.
The AWS DevOps Agent is available in Slack now. The direction is reversed here: AWS is putting its own agents where your engineers already are, so a DevOps question gets answered in the channel rather than in a console.
The Security, FinOps and Partner Central agents follow in the fall. If you already run employee agents in Slack, you now have a second vendor’s agents in the same workspace, and the governance question of who may invoke which agent gets a little wider.
What is dated for the fall
Two items are the interesting architecture, and both are Fall 2026.
The Amazon Connect and Agentforce Voice handoff is the one contact-centre teams should read twice. Salesforce describes it as Agent2Agent support between Agentforce Voice and Amazon Connect Customer, with real-time, bidirectional audio over WebSockets. The point is a live handoff rather than a cold transfer. A customer on an Amazon Connect call can be handed to an Agentforce Voice agent, or the reverse, with the conversation continuing instead of restarting.
For the many contact centres that run telephony on Connect and CRM on Salesforce, this is the first time the two vendors have described the seam as a protocol rather than a partner integration.
The handoff also changes the native contact centre decision. If Connect can hand a call to an Agentforce agent mid-conversation, the case for ripping out Connect to get Agentforce Voice gets weaker, and the case for keeping both and building the handoff gets stronger. I would not build on it until it ships, but I would stop planning a migration on the assumption it never will.
The A2A protocol underneath is the same one Agentforce uses to delegate to agents on Google and Azure.
The other dated item is the AWS Security, FinOps and Partner Central agents in Slack, also fall.
The Data 360 expansion is the part your data team will care about
Zero copy in Data 360 now reaches AWS Glue-managed Apache Iceberg tables, S3-backed Iceberg tables, Amazon Aurora, Amazon RDS and SageMaker Lakehouse. The existing Redshift and S3 support continues.
Read that list against what you have. If your lake is Iceberg on S3 with a Glue catalog, Data 360 can now federate it without the Redshift detour that some teams built to get there.
If your operational data is in Aurora or RDS, an agent can ground on it through Data 360 without a nightly copy. The file federation pattern I described in August covered Iceberg in general terms. This announcement names the AWS catalogs and stores it works with.
The usual cautions on zero copy still apply. Federation is a live query against a system you own, so the latency, the cost and the permission model of that system are now part of your agent’s behaviour. A slow Aurora query is a slow agent.
Informatica’s MCP servers are also expanding to general availability and are reachable from Amazon Bedrock AgentCore and from Quick. That gives an agent on AWS access to Informatica’s catalog discovery, data-quality scores and master-data retrieval. If your master data lives in Informatica, that is the path by which an AWS-side agent finds out which customer record is the golden one.
The procurement item
Salesforce Clouds are now in the AWS Marketplace across 33 countries, and customers can put AWS spend commitments toward Salesforce products, including Agentforce.
Not a feature, and it will not appear in a demo, but it may be the item that moves a deal. A company with a large committed AWS spend and a Salesforce renewal in the same year can now use one to satisfy the other. If you are the person who negotiates either contract, ask both vendors how the commitment counts before the next renewal, because the answer will not be in the press release.
How this sits next to the Google announcement
The two hyperscaler stories are different shapes, and it is worth being clear about which one you are in.
The Google announcement is about infrastructure. Hyperforce runs on Google Cloud, Gemini runs inside the reasoning engine, Salesforce shows up inside Gemini Enterprise. The story is about where Salesforce itself runs.
The AWS announcement is about adjacency. Salesforce stays where it is, and the connections to things you already run on AWS get thicker: models through Bedrock, data through zero copy, assistants through Quick, telephony through Connect. Nothing in it asks you to move Salesforce anywhere.
For most orgs that is the more practical story, because most orgs are not choosing a hyperscaler this year. They already have one, and the question is how much friction sits between it and their CRM. If you run Bedrock AgentCore for agents outside Salesforce, the same announcement makes the two agent estates easier to connect, through A2A on the voice side and MCP on the data side.
What I would do this quarter
If you are on AWS, four things, in order.
Ask your security team whether the Bedrock path changes the model approval you have. If Bedrock is already approved for other workloads, moving Agentforce inference there may shorten a review that was stuck on the Salesforce-hosted path. Check the Trust Layer settings still apply the way you expect, because the announcement says they do and you should confirm it.
Map your Data 360 sources against the new zero-copy list. Anything you currently copy from Aurora, RDS, or an Iceberg table on S3 into Data 360 is a candidate for federation instead. Measure the query cost before you switch, since federation trades storage credits for query credits, and the Flex Credit arithmetic changes with it.
If you run Amazon Connect, put the fall handoff on the roadmap and stop any migration plan that assumed you had to choose. Ask your Salesforce and AWS account teams for the preview timeline, because “Fall 2026” covers a lot of weeks.
If you use Quick, connect it and watch what people ask. The questions sellers put to Quick about their pipeline are the same ones they would put to Salesforce in Claude, and the permission model is the same. Whatever you learn about over-broad profiles from one surface applies to the other.
Understanding the basics
Can Agentforce use models on Amazon Bedrock?
Yes. As of September 15, 2026, Agentforce customers can access Anthropic and NVIDIA models through Amazon Bedrock, with OpenAI models listed as coming soon. Salesforce says the Bedrock path keeps zero data retention enforced and applies the same grounding and toxicity checks as the Salesforce-hosted path.
What is the Amazon Connect and Agentforce Voice handoff?
A planned Agent2Agent connection between Amazon Connect Customer and Agentforce Voice that carries real-time, two-way audio over WebSockets, so a call can move between the two without restarting. Salesforce has dated it for Fall 2026 and nothing has shipped yet.
Which AWS data sources does Data 360 zero copy now support?
AWS Glue-managed Apache Iceberg tables, S3-backed Iceberg tables, Amazon Aurora, Amazon RDS and SageMaker Lakehouse, in addition to the existing Redshift and S3 support. Federation runs live against the source, so the source’s performance and permissions become part of the agent’s behaviour.
Can I buy Salesforce through the AWS Marketplace?
Yes. Salesforce Clouds are listed in the AWS Marketplace in 33 countries, and AWS says customers can apply AWS spend commitments toward Salesforce products including Agentforce. How a commitment counts toward a specific contract is a question for both account teams.
The story that did not need a keynote
None of this is a new product. Seven existing things, on two clouds, with the seams between them made thinner. Less exciting than a reasoning model with a name, and for an org that already runs on AWS it is worth more, because every item on the list is something you can use with the systems you already have.
Start with the data list and the Bedrock approval. Both are decisions your own teams can make this month, and neither waits for the fall.
If you run Salesforce alongside AWS and want the zero-copy sources, the Bedrock model path, or the Connect handoff mapped against your own estate, talk to us, or start with the Agentforce readiness assessment.
