Twelve Agents. Half Work Alone.
The average company now runs 12 AI agents.
A Belitsoft industry report dropped last week with that number. Twelve. Up from near-zero two years ago. Projected to hit twenty by 2027.
The same report buried the stat that matters: 50% of those agents operate completely on their own.
Read that again. The average company runs a dozen AI agents — and half are running in isolation. No shared context. No coordination. No handoff. Each one executing its function in a silo while the other eleven do the same.
This is not a productivity stack. This is a fragmentation problem wearing a productivity costume.
What you actually built
Here’s the sequence that led here.
AI tools got good. Fast. A company added one for customer service. Another for analytics. One embedded in the developer stack. One the marketing team spun up without asking IT. A few individuals adopted personal tools because they didn’t want to wait for procurement.
Each addition was a local win. Each one did something the humans were doing before, faster and cheaper.
But the tools don’t know about each other. When the customer service agent flags a churn risk, the analytics agent doesn’t know. When the writing assistant drafts an outbound email, the CRM agent doesn’t see it. When someone resolves a decision in one thread, the tool running a parallel workflow on the same topic keeps going.
You didn’t build a coordinated stack. You built twelve smart silos.
The cognitive load plot twist
Here’s where this gets worse.
Research on AI and cognitive load is surfacing a counterintuitive result. High immersion in generative AI — over-reliance — intensifies the negative impact of cognitive strain rather than reducing it. People assumed they were offloading work. Some were. But the ones managing multiple AI systems were picking up new overhead they hadn’t priced in.
Because somebody has to coordinate. If the system doesn’t, the human does.
Every time you switch between tools, every time you re-paste context from one into another, every time you check whether the thing you asked one AI to do was caught by the other — that’s you doing coordination work. Manual. In your head. On borrowed bandwidth.
The promise was: AI reduces your cognitive load. The reality: twelve uncoordinated agents increases it. You swapped one category of tasks for a more fragmented version of the same ones. The open loops multiplied. The human became the routing layer.
What the 50% isolation stat is measuring
That “half work alone” number isn’t measuring capability. It’s measuring coordination debt.
Coordination debt is what you accumulate when you keep adding to a system without building the layer that holds it together. Technical debt slows your code. Decision debt clouds your strategy. Coordination debt fragments your execution. You keep adding pieces. Nothing connects. The value leaks through the gaps.
Companies have been accumulating agents at pace. They have not been accumulating the layer that makes those agents work together. The gap between “we run twelve agents” and “we run twelve coordinated agents” is the entire value proposition of the technology.
Dmitry Baraishuk, the CIO behind the Belitsoft report, said it plainly: “The winners will not be the companies with the most agents. They will be the ones that get their agents to work together and keep humans involved where it matters.”
That’s not a feature request. That’s a missing infrastructure layer.
The same pattern, smaller
This is easy to see at the enterprise scale. Twelve corporate AI deployments with no shared context shows up in audit failures, missed handoffs, contradictory outputs. It makes headlines.
The individual version is quieter.
You have your own fragmentation problem. It doesn’t look like a dozen enterprise deployments. It looks like four or five tools you’ve added over the past year — one for writing, one for research, one for meeting notes, one your team adopted, one you’re experimenting with. Each one is acting on your behalf. None of them know what the others are doing. None of them are tracking what the others have committed to.
You are the coordination layer. Every context switch, every re-paste, every manual check that nothing slipped — that’s you covering the gap.
The enterprise version of this problem is documented, funded, and beginning to get solved. Microsoft shipped governance infrastructure for it. Google and others will follow. The tooling exists to make enterprise agents coordinate.
The individual version is the same gap at a different layer. It’s still waiting.
The actual question
The question isn’t: how many AI agents should I have?
The question is: which of mine know about each other?
If the answer is none, you’ve accumulated coordination debt. And like all debt, it compounds quietly — until the missed handoffs, the re-opened loops, and the cognitive overhead of managing tools instead of using them start to feel like the normal cost of working with AI.
It’s not normal. It’s the cost of a missing layer.
Eliran Keren — Founder of Deeplica, building the coordination layer that makes AI tools work together instead of in parallel. The bottleneck isn’t intelligence. It’s coordination.
Sources: Belitsoft — 2026 AI Agent Trends Report · PMC — Effects of Generative AI on Cognitive Load · ACM CHI 2025 — AI and Critical Thinking in Knowledge Workers
