Salesforce’s Well-Architected team published a decision guide introducing a framework called “Orchestration Density” that tells architects and admins exactly when to use an AI agent, when to use traditional Flow automation, and when to use both. The guide includes a line that amounts to Salesforce telling its own customers to stop overusing Agentforce: “Using an agentic solution just to orchestrate and execute a deterministic path without any reasoning requirement is an anti-pattern.”
The Problem the Framework Solves
Agent sprawl is the term for what happens when teams deploy AI agents for every automation problem because agents are new and exciting. As Salesforce Ben’s analysis puts it, the side effect is “agent slop”: low-quality, unnecessary AI-generated output that adds noise without adding value. Too many overlapping agents become harder to govern, slower to audit, and often more expensive than a well-built Flow would have been.
The sprawl problem isn’t unique to Salesforce. Everest Group research across consumer goods firms, healthcare payers, and asset management firms identified agent sprawl vs. governance as one of seven unresolved tensions in enterprise AI. “Agent deployment decisions are often made at the team level, and the speed and distribution of adoption can quickly outpace centralized governance models,” the researchers wrote. “By the time oversight catches up, the agent estate is already too large and fragmented to govern retroactively.”
How Orchestration Density Works
The framework evaluates three factors in sequence, according to Salesforce’s official blog post on the framework:
Execution path is the first and most important factor. If every branch and outcome can be fully mapped at design time, traditional automation is the starting point. A workflow only needs an agent when the next step genuinely cannot be known until runtime because it depends on reasoning against live context, not just missing data.
Goal complexity asks whether each decision point can be expressed as an explicit rule. Salesforce is explicit that complexity alone doesn’t justify an agent. “Flow can handle enormously complex logic, hundreds of conditions, multiple objects, and sophisticated branching, and still be the right answer,” the blog post states. The threshold is whether decisions require judgment that can’t be reduced to a field value.
Modality mix identifies what kind of data flows in and out. Structured CRM fields are low modality. Email bodies or call transcripts require medium modality. Live audio, PDFs, or telemetry streams push into high modality. But Salesforce notes that high modality alone doesn’t mean high Orchestration Density: a single AI step called from an otherwise deterministic flow can be enough.
The Four-Level Matrix
The three factors combine into four architectural levels:
Low density: Rule-based, fully mappable, structured data only. Use Flow or Apex. Agents do not belong here.
Low-Medium density: Bulk records needing a single AI step like classification, scoring, or summarization. Use Agentforce Grid Batch Inferences.
Medium density: Rule-based overall, but some steps need AI reasoning. This is the hybrid zone: Flow or Apex handles orchestration, Agentforce handles reasoning-heavy steps, and Agent Script provides guided determinism where needed.
High density: Open-ended goals with both structured and unstructured inputs, competing outcomes, and decisions that emerge at runtime. This is where full Agentforce with Agent Script and MuleSoft Agent Fabric for multi-agent collaboration belongs.
The key insight, according to Salesforce Ben, is the middle two rows. Most teams think the choice is binary: Flow or agent. The framework argues that a large share of real-world use cases live in hybrid territory, where Flow handles the specifiable parts and agents handle only the reasoning-intensive steps.
Where This Fits in the Enterprise AI Maturity Cycle
Salesforce publishing a framework that explicitly tells customers not to overuse its newest product is a signal worth noting. It follows a pattern across the enterprise AI market where the initial “deploy agents everywhere” enthusiasm is giving way to cost-conscious, fit-for-purpose deployment models.
The Orchestration Density guide includes worked examples showing agent patterns that look impressive in demos but fail the framework’s criteria. An outage notification system powered by a “swarm of autonomous agents” scores low density on all three factors because triggers are binary, decisions are rule-resolvable, and the data is structured. Quote pricing with an AI prompt template also fails: if the rules can be written in prose, they can be written in logic.
The full decision guide includes additional product comparisons and use-case mapping tables beyond the density framework.