Salesforce takes Agentforce multi-agent orchestration to GA — Atlas Reasoning Engine 3.0, A2A, and hosted MCP servers ship in Summer 26

Salesforce announced on June 15, 2026 that Agentforce multi-agent orchestration has reached general availability as part of the Summer '26 product release, with the rollout staged in waves starting June 13. The capability moves from beta to GA, and the release ships the supporting reasoning, protocol, and tooling layers alongside it. Salesforce's own Summer '26 release page was bot-blocked (HTTP 403) when we tried to fetch it directly, so the facts below are drawn from multiple reachable secondary sources that independently describe the same GA date, components, and scope.

What multi-agent orchestration actually does. Instead of a single agent handling an entire task, orchestration lets a single orchestrator agent coordinate a set of specialist subagents. The orchestrator inspects which subagents are registered, reads their descriptions and available actions, and routes each piece of work to the specialist best suited to it — dynamically, rather than through fixed decision trees. The practical effect is that complex, end-to-end workflows can be split across purpose-built agents while the customer or user still deals with a single point of contact that carries shared context across channels.

Atlas Reasoning Engine 3.0 is the coordination layer. The GA release ships Atlas Reasoning Engine 3.0 as the engine powering the routing and reasoning behind orchestration. This is the component that reads subagent descriptions and decides where work goes, which is why Salesforce frames agent descriptions as load-bearing routing input rather than user-facing documentation.

A2A and hosted MCP servers extend the reach beyond Salesforce. The release adds Agent2Agent (A2A) support, letting customers securely connect Agentforce to third-party agents — the goal being an agentic enterprise that reaches beyond the Salesforce platform itself. Note the scope limit: advanced cross-platform A2A beyond Salesforce Cloud is described as still in beta, so the fully open cross-vendor path is not yet GA. Separately, Salesforce-hosted MCP servers reach GA in this release, exposing capabilities such as SObject CRUD and SOQL, Data 360 queries, and Tableau analytics as tool sources that agents can call.

The GA label does not mean deploy-on-day-one. Because routing depends entirely on the quality of each subagent's description and registered actions, a multi-agent setup that is misconfigured will route work to the wrong specialist. Operators evaluating this should treat the description and action metadata as the real work — get single agents and their descriptions right first, then layer orchestration on top once the routing inputs are trustworthy. Rushing several subagents into an orchestrator before their descriptions are precise is the predictable failure mode.

Where this sits in the market. This is Salesforce moving its agent platform from "one agent per task" toward coordinated teams of agents with a cross-vendor protocol story — the same direction Microsoft (Copilot Studio agent-to-agent), Google (managed agents, A2A), and others have been pushing in 2026. For teams already invested in Salesforce, orchestration GA is the signal that multi-agent workflows are now a supported production path rather than a preview, with the caveat that the cross-platform pieces are still maturing.