App sprawl was something you could survive by ignoring. Agent sprawl bills you monthly, holds credentials, and now degrades the thing meant to fix it.
Alcon, a regulated medical device company, ended up with more than nine hundred AI agents built in silos across the organization in under a year. Their own framing of the problem, in public, put security ahead of cost and ahead of duplication. “It’s a security risk, first and foremost.”
Nine hundred. Not pilots in a sandbox. Built, in production, inside a company that operates under FDA oversight and knew better.
The reaction to that number is usually that Alcon must have done something unusual. They did not. They moved slightly faster than everyone else down a road that the entire ecosystem is currently paving, and they were honest about where they ended up. What makes the story useful is not that it is extreme. It is that it is early.
The benchmark everyone quoted and nobody acted on
Salesforce published its 2026 Connectivity Benchmark in February, the eleventh annual edition, surveying 1,050 enterprise IT leaders with Vanson Bourne and Deloitte Digital. The headline numbers got wide pickup:
- The average enterprise runs 12 agents, projected to grow 67% within two years.
- 83% of organizations report that most or all teams and functions have adopted agents.
- 50% of agents operate in isolated silos rather than as part of a multi-agent system, which Salesforce’s own writeup connects to disconnected workflows, redundant automations, and shadow AI risk.
- Only 54% have a centralized governance framework for agents.
- 86% of IT leaders are concerned that agents will introduce more complexity than value without proper integration.
That last figure deserves a second read, because it is a vendor publishing survey data in which a large majority of its own buyers say they are worried the thing they are buying will make their environment worse. Salesforce published it anyway, which is to its credit, and then shipped four more agents at Connections in June.
The number that got the least attention is the most damning. Only 11% of the prior year’s planned agentic use cases reached production. Set that against 12 agents already running, and the picture is not an organization executing a strategy. It is an organization accumulating things.
Sitting underneath all of it: enterprise application count rose from 897 to 957 year over year, with 27% integrated. Agent sprawl is not replacing app sprawl. It is being layered on top of it, at a faster rate, on the same unintegrated foundation.
Why this is not just app sprawl again
The instinct is to file this as a familiar problem. Organizations accumulated 957 applications; now they will accumulate agents; IT will eventually rationalize the portfolio the way it always has. Annoying, survivable, cyclical.
That instinct is wrong on three counts, and each one is structural.
Agents have a meter. A dormant application costs a license, and unused licenses are the easiest line item in the world to find and cut. An agent under Flex Credits costs roughly $0.10 per standard action and $0.15 per voice action, drawn from credits purchased at $500 per 100,000. The published rate card is explicit that credits must be consumed before the order end date, with no rollover. Pre-production consumption is real: sandboxes and scratch orgs draw credits too, at a reduced but nonzero rate. So a redundant agent is not idle overhead. It is a variable cost with no cap, attached to something nobody owns, that only shows up in aggregate on a bill somebody in finance is trying to reconcile against a forecast built when there were four agents instead of twelve.
Agents have an identity. An unused application sits there. An agent reads data, calls actions, and writes records under some permission context, at machine speed, without an interactive login. Every agent is a non-human identity with a scope, and the scope was set by whoever built it, usually to whatever made the agent work during testing. This is precisely the class of identity that the OAuth supply chain attacks of the last twelve months exploited, and it is precisely the class that gets monitored least. Alcon leading with security rather than cost is not a regulated-industry quirk. It is the correct ordering.
Agents now compete with each other. This is the new one, and it is the reason the timing matters.
The orchestration trap
Multi-Agent Orchestration reached general availability on June 15, 2026, as the headline feature of Summer ’26. It is positioned as the cure for the silo problem: a primary orchestrator agent receives a request, routes it to the best-fit specialist, and holds shared context so a customer moving from chat to email does not repeat themselves. Atlas Reasoning Engine 3.0 sits underneath it, with MCP for tool access and A2A for coordination with agents on other platforms.
The mechanism is the part worth understanding. Atlas does not route through a decision tree or a rules table. It reads each registered subagent’s natural language description, instructions, and available actions, and decides from that which specialist is best equipped for the task. Routing accuracy is a function of description quality rather than rule configuration.
Sit with the implication.
In a single-agent world, a badly described agent is a badly described agent. It affects itself. In an orchestrated world, every agent’s description is an input to a live routing decision, which means agents that overlap in purpose or describe themselves vaguely do not just underperform individually. They actively degrade the routing of the agents around them. Twelve well-scoped agents with precise, mutually exclusive descriptions orchestrate well. Twelve agents built independently by four teams, half of them siloed, with descriptions written as internal documentation or as marketing copy, produce a routing surface where the orchestrator’s job is genuinely ambiguous.
Agent count stopped being purely a cost and governance problem the day orchestration went GA. It became a quality problem, and it scales in the wrong direction. The organization heading for twenty agents by 2027 is not just accumulating spend and identities. It is accumulating routing ambiguity, and the fix that was sold to solve sprawl gets less reliable as sprawl increases.
There is a second-order version of this at the seams. When work passes between agents, the handoff is where context is lost and where ownership of a failed workflow becomes genuinely unclear. If a customer gets a wrong answer after passing through a triage agent, a billing agent, and an entitlements agent, the question of which agent was wrong is a real investigation, not a lookup.
What good looks like
Alcon’s response is the useful template, and it is notable for what it is not. They did not buy a governance product first. They established a formal AI governance board, enforced human-in-the-loop controls for customer-facing agents, and standardized access through API-led architecture they had already invested in years earlier, exposing existing MuleSoft APIs so agents across Agentforce, Bedrock, and Azure could reuse them safely rather than each building its own path to the same data.
Sequence, not tooling. Four things, in order.
Inventory before consolidation. Every agent, its owner, its business justification, its permission scope, its action list, and its monthly credit consumption. Most organizations cannot produce this today. Producing it is usually where the first round of decisions makes itself, because a meaningful fraction of the list will have no owner, no justification anyone can articulate, or an obvious duplicate two departments over.
Consolidate on purpose, not by count. The goal is not fewer agents as a target number. It is agents with non-overlapping domains. Two agents that both plausibly handle a billing question are worse than one agent that handles billing, even if the two are individually better built. Overlap is the specific thing that poisons routing.
Treat descriptions as production artifacts. An agent description is not documentation and it is not a listing blurb. It is a routing input evaluated by a reasoning engine on every request. It should state explicitly what the agent handles and, just as importantly, what it does not. It belongs under version control and in code review, held to the standard of an API contract, because functionally that is what it now is.
Govern before the next wave, not after. 54% having a governance framework at 12 agents means roughly half the market will hit 20 agents with no framework at all. The window where this is a tidy afternoon of work is open now and closes on its own.
The uncomfortable question worth putting to any agent roadmap: for each agent proposed, what would break if this were a topic inside an existing agent instead of a new one? The answer is often nothing, and the agent gets built anyway, because building a new agent is the path of least organizational resistance. It requires no negotiation with another team’s roadmap and no shared ownership. That incentive is what produced nine hundred agents at Alcon, and it has not changed at anyone else.
Resource Interactive designs agent architectures around domain boundaries rather than workflow count, on the theory that the number of agents an organization needs is almost always smaller than the number it will otherwise build.
Sources
- Salesforce, “Salesforce Announces 2026 Connectivity Report,” February 2026
- MuleSoft Blog, “Agentic Transformation Reaches New Heights: 2026 Connectivity Benchmark Report Insights,” February 2026
- SalesforceDevops.net, “From Integration to Intelligence: What Salesforce’s 2026 Connectivity Benchmark Really Signals,” February 2026
- Salesforce Ben, “How Many AI Agents Is Too Many? Salesforce Adds Four More at Connections,” June 2026
- Salesforce, “Agentforce Multi-Agent Orchestration” product documentation
- Salesforce, Summer ’26 Product Release Announcement, May 2026
- Salesforce Flex Credits Rate Card, April 21, 2026
- Enterprise DNA, analysis of the 2026 Connectivity Benchmark, April 2026