Agentforce in Summer '26: New Builder, Agent Script, MCP, and Multi-Agent Orchestration

If Summer '26 has a beating heart, it's Agentforce. Salesforce shipped a wave of platform updates that change how you build, govern, and run AI agents — and a few of them are big enough to change your roadmap. Here's what matters, in plain terms.

The new Agentforce Builder is now the default path

Salesforce is steadily retiring the legacy agent-building experience. Starting the week of July 13, 2026, the "New Agent" button disappears from the legacy Agentforce Builder in Setup. From that point on, you create agents only in the new Agentforce Builder inside Agentforce Studio.

The good news for anyone with existing agents:

• Your legacy agents keep working. You can still edit, activate, and version them.
• You can upgrade a legacy agent to the new builder in just a few clicks. Previously this meant manually rebuilding the agent. Now Salesforce creates a new version for you — carrying over subagents, actions, system messages, settings, data, and connections, all converted to Agent Script.
• After upgrading, Agentforce can optimize the agent for Agent Script to improve reliability and performance.

If you've been putting off the migration, this is your nudge. The path is now low-effort, and the deadline is real.

Agent Script: deterministic logic meets LLM reasoning

The new builder is powered by Agent Script, a way to build agents that blend deterministic, rule-based logic with the reasoning power of large language models. This "hybrid reasoning" approach is the core idea: you get the predictability of defined steps where you need it, and the flexibility of an LLM where you don't.

Summer '26 adds a steady stream of Agent Script capabilities, including ending a session with utils.end_session, creating variable lists of objects, accessing voice call properties through linked variables, and nesting And/Or operators with parentheses for more precise logic.

You can now choose your model — including Gemini and Claude

One of the most requested capabilities lands here. You can pick the LLM that powers your agents:

• Salesforce Default — uses OpenAI.
• AWS-Hosted — uses Anthropic (note: the AWS-hosted option is moving to Claude Haiku 4.5 on Amazon Bedrock, and Claude Opus 4.6 is generally available on the Einstein Platform).
• Google Gemini — uses Gemini 3.5 Flash, available for agents built in the new Agentforce Builder only.

A practical heads-up: if you switch models, re-test your prompts, custom actions, and topics. Different models behave differently, and a prompt tuned for one can drift on another. For consistency, update custom actions to match the model you've chosen.

MCP for Agentforce: open standards, less integration grunt work

This is a genuinely strategic addition. Model Context Protocol (MCP) is an open standard for connecting AI models to external systems, and Salesforce has built an MCP client directly into Agentforce.

In practice, that means the painful parts of integration — API connectivity, authentication, and data access — are handled for you. You can register and manage MCP servers, give agents secure access to MCP server tools, and govern their behavior with Agentforce Policies that control access, set usage limits, and enforce governance standards.

If you've been wiring up custom integrations for every external system your agents touch, MCP is the shortcut you've been waiting for.

Multi-Agent Orchestration (Beta): agents that call other agents

With Multi-Agent Orchestration, your Agentforce agent can connect to other specialized agents in your org and collaborate on complex tasks. Instead of forcing a user to juggle multiple disconnected sessions, one agent acts as a single point of contact and delegates to connected subagents behind the scenes.

This is the building block for sophisticated, layered automations — a "generalist" agent that hands off to specialists for billing, scheduling, or research.

Voice agents get sharper and more global

Agentforce Voice picked up meaningful quality and reach improvements:

• Sharper transcription and more natural speech — better accuracy in noisy environments, with accented speech, and with specialized terminology. No configuration or restart required.
• Custom pronunciations for non-English personas using IPA or CMU notation.
• Voice-enabled agents in the new Agentforce Builder, including for Enhanced Chat v2, so customers can switch between text and voice mid-conversation.
• SIP routing for voice calls, in addition to traditional PSTN phone numbers.
• Expanded global language support (Beta).

Better observability and governance

You can't trust what you can't measure. Agentforce Observability now offers Refined Agent Analytics — a unified dashboard combining Service Agent Analytics and Employee Agent Analytics, with more than 40 metrics across Quality, Health, Effectiveness, and Usage. You can drill from high-level views into subagents, intents, and actions, and use Custom Scorers (Beta) to evaluate sessions against your own KPIs.

On the cost side, the new AiAgentGenerativeAiUsage data model surfaces your Agentforce and generative AI consumption (metered and unmetered), updated every five minutes, so you can build reports and dashboards on AI spend.

Prompt Builder upgrades worth knowing

Prompt Builder also leveled up:

• Block Editor (Beta) for authoring prompts faster.
• Conditional logic (Beta) to build smarter, more dynamic prompt templates.
• Keyboard shortcuts for power users.
• View prompt template dependencies across your org.
• Deploy and version-control prompt templates with the Metadata API — finally treating prompts like the deployable assets they are.

Frequently asked questions

Do I have to migrate my agents to the new Agentforce Builder? You should. Starting the week of July 13, 2026, you can only create new agents in the new builder. Existing agents still work in the legacy builder, but the upgrade is now a few clicks and carries everything over.

Can I use Claude or Gemini instead of OpenAI in Agentforce? Yes. You can select Salesforce Default (OpenAI), AWS-Hosted (Anthropic Claude), or Google Gemini (Gemini 3.5 Flash). Gemini requires agents built in the new Agentforce Builder.

What is MCP for Agentforce? It's a built-in Model Context Protocol client that connects your agents to external systems using an open standard, handling API connectivity, authentication, and data access for you.

What is Agent Script? A way to build agents that combine deterministic logic with LLM reasoning, giving you both predictability and flexibility in the same agent.

Can one agent use another agent? Yes — Multi-Agent Orchestration (Beta) lets an agent connect to and coordinate other specialized agents in your org.

The bottom line

Summer '26 turns Agentforce from a promising feature set into a real platform: a modern builder, your choice of model, open-standard integrations via MCP, agents that collaborate, and the observability to keep it all accountable. If you build on Salesforce, this is the area to invest your time in now.

Next: see how agents show up in your day-to-day automations in our Flow and Automation guide.