Salesforce Development in the Age of Agentforce: What Changes for Engineering Teams

Salesforce says its AI agents have now completed 7.0 billion discrete tasks across Agentforce and Slack, 3.2 billion of them in the quarter that ended on July 31, 2026. Agentforce annual recurring revenue passed $1.5 billion in the same quarter, though the company widened the metric to include Slackbot and Headless 360, so the reported growth of more than 240% is not entirely like for like. Either way, agents have left the keynote stage and entered production orgs. That changes what a Salesforce engineering team spends its week on. Less of the job is building screens for people to click through. More of it is building the actions, data and guardrails an agent calls on its own, and getting that layer right is part of why Crunch-IS is a leader in Salesforce development services for companies moving Agentforce past the pilot. Here is what shifts in the backlog.

Readiness lags enthusiasm

Salesforce’s fourth State of IT report surveyed more than 4,000 IT leaders, 2,173 of whom work in or lead software development. 82% said they use or plan to use agents within two years. Yet fewer than half of the developers felt fully confident they understood how agents work, and 68% said their teams lacked the resources to build and deploy agents at scale. Only 53% of IT leaders fully trusted the accuracy of their own data.

That last figure is the one that matters. An agent reasons over whatever the org contains.

Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. It also expects a third of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024. Both forecasts can hold at once. Lots of agents get built, and many of them get switched off.

Five shifts in the engineering work

1. Apex gets a non-human caller. Agents act through invocable Apex, flows and prompt templates. An invocable method used to have a predictable caller. Now a language model decides when to call it and what to pass in, so input validation, precise descriptions and idempotent behaviour become requirements. Governor limits do not relax for agents: 100 SOQL queries and 10,000 milliseconds of CPU per synchronous transaction apply whoever triggered the action.
2. Descriptions become code. The label and description on an action are what the planner reads when it picks one. A vague description sends the wrong request to the wrong action, and no compiler will flag it. Salesforce’s vocabulary is moving the same way; since April 2026 its documentation calls agent topics “subagents”.
3. Testing moves from assertions to evaluations. A unit test checks that a method returns the expected value. An agent test checks that, across dozens of phrasings of the same request, the agent picks the right action and leaves the wrong ones alone. Agentforce DX, generally available since May 2025, runs these test specs from the Salesforce CLI, which puts them in the same CI pipeline as Apex tests.
4. Integration becomes the critical path. MuleSoft’s 2025 Connectivity Benchmark found the average enterprise runs 897 applications, with only 29% of them connected. 95% of the 1,050 IT leaders surveyed said they struggle to integrate data across systems. An agent that can only see Salesforce answers from Salesforce’s partial view of the customer.
5. The data model gets read by a machine. Data 360 ingested 104 trillion records in Salesforce’s latest quarter, 82 trillion of them through Zero Copy. Every duplicate field, stale picklist value and ambiguous relationship in that data is now something an agent can misread with total confidence.

What Salesforce’s own service desk shows

Salesforce has been its own first customer. Marc Benioff said in 2025 that support headcount had fallen from about 9,000 to about 5,000 as agents took on conversations, which he described as roughly half handled by agents and half by humans. Whatever one makes of the staffing decision, the engineering lesson is in the human half. Every case an agent cannot close has to reach a person with the history, the customer’s words and the reason for escalation intact. Escalation design is engineering work. It is also the part most pilots leave until last.

Who does what on the team

The roles are drifting. Admins who used to build flows now write action instructions and guardrails. Developers spend more time on evaluation suites and observability, meaning logs of which action an agent chose, with which inputs, and why. Architects get a new standing question in every design review: what should an agent never be allowed to do, and where does that rule live?

Permission sets matter more than they did. An agent runs under a user context and can do anything that user can.

The constraint that will decide most projects

The agent layer is the new part. The data and integrations underneath it are old, and that is where most of Gartner’s cancelled projects will come from. Teams that spend their first quarter fixing the data model, connecting the three systems the agent actually needs and writing evaluations before prompts will ship agents that stay switched on. Teams that start with the demo will get an agent that does the wrong thing quickly, politely and at scale, inside a production org.