Twelve Agents, Half of Them Working Alone

The average company now runs twelve AI agents, and half of them operate in complete isolation from the others.

The bottleneck stopped being how many agents you can build. It became whether any of them can hand work to another.

The fix is boring and unglamorous: a shared context layer and a human who owns the handoffs, not a thirteenth agent.

Salesforce put a number on something I have been watching go wrong in client deployments for a year. Their 2026 Connectivity Benchmark Report found the average company runs twelve AI agents, expected to hit twenty by 2027, and that half of those agents work entirely on their own. They do their one job, return a result, and have no idea the other eleven exist.

I read that and thought about a logistics company I sat with in the spring. They were proud of their agent count. Sales had a lead-qualification agent. Support had a triage agent. The ops team had built something that reconciled shipment records overnight. Finance had a contract-reading agent a vendor sold them. Five agents, five teams, five separate wins on five separate slide decks. And not one of them could pass anything to another.

What working alone actually looks like

When an agent works alone, the seams show up as human labor.

The support triage agent at that logistics company flagged a billing dispute. Good. Then a person copied the dispute into an email, forwarded it to finance, where someone pasted the relevant lines into the contract agent, read the output, and typed a summary back to support. Two agents that should have been one workflow were stitched together by a junior employee acting as a USB cable.

This is the pattern in almost every deployment I see. The agents are fine. They do their narrow task well. But the moment a real process crosses a team boundary, and almost every process worth automating does, the agent hits a wall and a human carries the work over it by hand. You did not remove the manual step. You moved it to the gap between agents and stopped measuring it.

The Salesforce number is generous, honestly. Twelve agents that half-work in isolation is not twelve units of automation. It is six, plus a new category of invisible glue work that nobody put on a dashboard.

Why it happens every time

The isolation is not a technical accident. It is the org chart leaking into the architecture.

Each team buys or builds the agent that solves its own pain. Sales does not wait for a platform team. Support does not file a ticket and queue for two quarters. They each grab a tool, wire it to their own data, and ship. Twelve months later you have twelve agents that reflect twelve procurement decisions and zero shared design.

None of them share context because nobody owned the context. The lead-qualification agent knows the prospect. The support agent knows the same company as a ticket number. The finance agent knows them as a contract ID. Three agents, three names for one customer, no connective tissue between them. The agent cannot hand off because it has nothing to hand off to and no shared language to do it in.

This is the same failure I keep writing about from a different angle. People treat the agent as the hard part and the surrounding system as an afterthought. The model was never the constraint. The plumbing is.

The connective tissue exists now

The encouraging part is that the missing layer is finally real infrastructure instead of a slide in a vendor pitch.

Anthropic's Model Context Protocol crossed ninety-seven million installs this spring. That matters because MCP is precisely the shared language those three agents were missing: a standard way for an agent to reach the same tools, the same records, the same context that another agent already uses. When your support agent and your finance agent both speak MCP to the same customer system, the handoff stops being an email a human sends.

The orchestration piece arrived too. Anthropic shipped multiagent orchestration and managed-agent webhooks on its platform in June. That gives you a supervisor that can route a task across several agents and wait for each to finish, instead of a person watching a queue and forwarding outputs. The building blocks for connected agents are no longer the bottleneck. The decision to use them is.

The metric that actually moves

Here is what I tell teams who show me their agent count like it is a scoreboard. The number of agents you run is not a measure of anything. It is an inventory, and inventory is a cost, not a result.

The question that matters is whether a unit of real work, a billing dispute, a qualified lead that becomes a signed contract, a shipment exception, can move from start to finish without a human carrying it across a gap between two agents. Count the handoffs that still require a person to copy, paste, summarize, and forward. That number is your actual automation debt, and it is usually far higher than anyone wants to admit.

This connects to the framework I keep coming back to. Someone has to set the bar, define what done means for the whole workflow and not just one agent's slice of it, and someone has to hold the bar, verify the result end to end. When agents work alone, nobody is doing either job at the workflow level. Each agent has a tiny local bar and the space between them is unowned. That unowned space is where the work quietly piles back onto your team.

What to do this quarter

Do not build a thirteenth agent. You have enough.

Pick one process that already crosses two of your isolated agents and that costs real human time in the seam. Map the handoff. Then wire those two agents to a shared context layer so the output of the first becomes the input of the second without a person in between. One connected workflow that runs clean is worth more than five proud agents that each need a babysitter at the border.

The companies that win the next two years will not be the ones with twenty agents by 2027. They will be the ones whose agents can hand work to each other, with a human holding the bar where it actually matters. Twelve agents working alone is not a head start. It is twelve walls you now have to knock down.