Salesforce Summer '26: The First-90-Days Agentforce Plan

The Salesforce Summer '26 release is the largest Agentforce update yet, and it graduates multi-agent orchestration from beta to general availability on June 15, 2026. Most coverage stops at listing the ten headline features. The harder question — the one every RevOps and IT leader actually has to answer — is what to turn on first, and in what order, to get value without breaking production.

That sequencing matters because Summer '26 is not one feature; it is roughly ten headline innovations plus seven additional updates spread across twenty-three product pillars, documented in a release notes set that runs to several hundred pages (a vendor-stated figure of 822 pages). Deploy everything at once and you create work and risk with no clear payoff. Deploy in the wrong order and you stand up multi-agent orchestration before your agents are described well enough to route reliably.

This guide is a prioritized 90-day deployment plan. It starts with the two features that move pipeline fastest with the least friction, sequences multi-agent orchestration into months two and three for a reason, and flags the API v67.0 security change you should audit before — not after — your org upgrades. Every fact, number, and quote below is sourced to a primary release document or named analyst.

Summer '26 rolls out in production waves starting June 13, with general availability on June 15. Sandbox preview opened May 8, and the first production wave landed May 15. The release is organized around a single idea: the agentic enterprise stops being a marketing story and becomes how the platform works. The pieces that matter most for a deployment plan fall into three groups — agents that act, features that feed agents context, and the orchestration layer that coordinates them.

The two features to know on Day 1 are the Customer Engagement Agent and Momentum. The Customer Engagement Agent is a net-new capability that independently interacts with and qualifies buyers around the clock across website and email, then hands warm leads to sales — addressing the perennial problem that good leads go cold because reps cannot follow up fast enough. Momentum, the conversation-intelligence product Salesforce announced it was acquiring on February 18, 2026, captures and structures every customer interaction — calls, emails, and meetings — and writes that data back to Salesforce in real time, connecting Zoom and Google Meet to Agentforce context.

Beyond the marquee agents, Summer '26 quietly ships the plumbing that makes agents useful: Salesforce-hosted MCP servers reach GA, exposing SObject CRUD and SOQL, Data 360 queries, and Tableau analytics as tool sources that external AI agents can consume, while Tableau MCP makes your analytical data a first-class source for agent reasoning. For teams already building on Agentforce, our Agentforce CRM automation guide covers the foundation these upgrades build on.

One caveat first, because it is easy to get wrong: the headline Agentforce revenue figures are not Summer '26 launch numbers. Agentforce ARR reached $800 million as of Salesforce's Q4 FY26 earnings — the quarter ended January 31, 2026, reported in February 2026, several months before this release. The right way to read them is as context: this is the adoption momentum that makes Summer '26's GA features consequential rather than incremental. Every figure here is vendor-stated from Salesforce's own earnings disclosures.

With that framing, the trajectory is striking. Agentforce ARR grew 169% year over year and 48% sequentially, from $540M in Q3 FY26 to $800M in Q4. Combined Agentforce and Data 360 ARR exceeded $2.9 billion, up over 200% year over year. Salesforce reported more than 29,000 cumulative Agentforce deals closed since launch — up roughly 50% quarter over quarter — and more than 60% of Agentforce and Data 360 bookings in the quarter came from existing-customer expansion, not new logos. That last detail matters for a deployment plan: it signals customers who already run Agentforce are buying deeper, which is exactly the expansion path this 90-day sequence is built for.

Here is the proprietary view no other Summer '26 piece publishes: a single matrix that sequences each major feature by phase and pairs it with what you need in place first, its expected pipeline impact, its deployment complexity, and whether it is GA or beta in this release. Read it top to bottom as a deployment order — the rows are already prioritized. Impact and complexity ratings are our own assessment based on the release documentation and independent technical analysis; treat them as planning guidance, not vendor-certified scores.

The shape of the plan is deliberate. Days 1-30 are reserved for features with high or medium pipeline impact and low deployment complexity that require no change to how reps work. Days 31-60 take on multi-agent orchestration, the highest-impact-but-highest- complexity item, only after the prerequisite description work is done. Days 61-90 add surface-area features for the orgs ready to extend further. Storefront Next is marked "varies" because its impact depends entirely on whether you run Commerce Cloud — for non-commerce orgs it does not belong in the plan at all.

The first month is about velocity, not ambition. Stand up the Customer Engagement Agent and connect Momentum, because both deliver measurable pipeline value while requiring almost nothing from your sales team in terms of behavior change. The Customer Engagement Agent handles the part of the funnel humans are worst at — instant, consistent follow-up at any hour — and only escalates a warm, qualified lead to a rep. Momentum solves the inverse problem: it makes sure the conversations reps are already having actually land in the CRM as structured data, instead of evaporating into notes that never get written.

Pair them with Collections with Agentforce if you have a finance team chasing invoices. It scores invoices using payment history and account data, then recommends risk-based dunning plans — outreach cadence, escalation timing, and when to apply fees or service holds — with dynamic rules by customer tier. It is low-complexity because the data it needs (invoices and payment history) already lives in your org, and it produces a clear, near-term cash-flow benefit.

Multi-agent orchestration is the headline of Summer '26, and it is genuinely powerful — but it is the one feature you should not rush. The orchestration is driven by the Atlas Reasoning Engine 3.0: a coordination layer that understands intent, determines the data and actions a task needs, identifies which agent or tool is responsible, and executes. Critically, the orchestrator routes by reading the descriptions of registered subagents — there are no fixed decision trees. That design choice is exactly why description quality, not deployment speed, is the constraint.

In practice, an agent's description functions as a load-bearing routing input, not as user documentation. If your subagents are described vaguely, the orchestrator routes badly, and you get the worst of both worlds: more agents, less reliability. So the first half of this phase is unglamorous and essential — write precise, action-accurate descriptions for every agent you intend to register, then test routing with a small set before expanding the roster.

The second risk to plan for in this phase is what independent analysts call the "seam problem." Multi-agent systems create failure modes at every handoff point — incorrect orchestrator routing, specialists acting on stale data, context lost in a handoff, and latency between agents causing timeouts — and the count of those failure surfaces grows much faster than the number of agents. The ChatForest technical analysis frames it as roughly an N-squared growth in failure modes rather than a linear one; that figure is an analyst observation, not a Salesforce-stated fact, but the directional point is sound and worth designing against.

Atlas 3.0's answer is structured context packets that travel with every delegation, preserving context across handoffs so a subagent does not act on a stale or truncated picture. The same release adds the Agent-to-Agent (A2A) protocol, which lets Agentforce communicate with agents on Google Vertex, Microsoft Azure Agent Mesh, and third-party frameworks via capability manifests and authenticated, cryptographically-identified delegation. Note that advanced cross-platform A2A beyond Salesforce Cloud remains in beta — so for a Days 31-60 plan, keep orchestration within your own agent estate first and treat cross-platform routing as a later experiment. The five multi-agent patterns that translate to live revenue are a useful design reference before you register your first orchestrator.

The final phase is opt-in by org type. Slack First Sales brings Agentforce Sales into Slack so sellers can prospect, engage leads, and manage pipeline without context-switching to the CRM — the vendor-stated goal being to let a single seller operate with a full revenue team behind them and scale without adding headcount. It is a Days 61-90 item rather than Day 1 because the value compounds once your sales motion is already running inside Slack and your orchestration layer is stable.

Storefront Next is for Commerce Cloud merchants: an AI-first, enterprise-grade storefront designed to deliver fast, out-of-the-box conversion without heavy developer resources. If you do not run commerce, skip it entirely. The IT Service Domain Pack, with its 50+ pre-built agents, is an internal-productivity play rather than a pipeline one — worth scheduling here for orgs that want agentic IT support across Slack, Teams, and the service desk. As always, sequence these against your own readiness, not the calendar; the point of the plan is order and prerequisites, not a rigid clock.

Before any of the above, there is a step that has nothing to do with agents and everything to do with not breaking production. API v67.0 ships with Summer '26 and changes a security default: SOQL, SOSL, DML, and Database methods now default to user mode — enforcing object permissions, field-level security, and sharing rules — instead of system mode, and Apex classes default to with sharing. This is a sensible security hardening, but it means custom code written for earlier API versions can behave differently once your org upgrades.

The practical implication: if you have legacy Apex that quietly relied on system-mode access to read or write records a running user could not otherwise see, that code may start returning fewer rows, failing on field access, or throwing where it used to succeed. Because the change is buried in the release notes rather than headed up as a marquee feature, it is easy to miss — and the failure mode is subtle, not loud. Run a code audit on your custom Apex before your org reaches GA, not after, so you find these regressions in a sandbox instead of in production.

Step back from the feature list and a strategy comes into focus. Salesforce is not merely deploying point AI features; it is building two layers in parallel — an orchestration layer in Slack and Agentforce, and a context layer in Data Cloud plus conversation ingestion. That is the lens through which the Momentum acquisition makes sense. Salesforce did not buy a meeting-notes tool; it bought a way to feed agents the unstructured conversation data that determines whether their reasoning is grounded or guessing. As Salesforce CPO Steve Fisher put it, delivering on agents' promise requires visibility and context from every meaningful interaction.

The forward-looking read is that the competitive battleground for CRM AI is shifting from feature parity to context fidelity. Anyone can announce an agent; the durable advantage goes to whoever can ground that agent in accurate, current data about the customer. That is why this plan front-loads the unglamorous work — connecting conversation capture, writing precise agent descriptions, auditing data access — rather than the demos. For teams weighing platforms before committing to this sequence, our comparisons of how Salesforce's Agentforce pricing compares to HubSpot AI and the broader CRM AI agent landscape across Salesforce, HubSpot, and Zoho put Summer '26 in competitive context.

The practical conclusion for an implementation team is the same one that runs through this whole plan. Treat Summer '26 not as a shopping list to clear in a sprint, but as a staged capability buildout where the lowest-friction agents earn trust first, the orchestration layer is earned through description and data quality, and the context plumbing is the real long-term investment. This is exactly the kind of sequencing we build into CRM automation engagements and broader AI transformation programs.