Everyone Ships AI Features. Almost No One Ships AI Teams.
The actual shift ahead.
A single-purpose assistant is a tool. What’s becoming possible - and what a handful of serious companies are quietly building- is something different: a set of specialized agents that divide work between themselves, negotiate handoffs, and move a goal forward without a human steering every step. Call it a multi-agent system. Call it a digital team. The framing matters less than what it unlocks: the work that today sits in the cracks between departments, systems, and spreadsheets finally getting done. Let me make that concrete.
A sample scenario:
A regional water utility. 3 a.m. A pump in a pressure station drifts - vibration up, bearing temperature climbing. Nothing dramatic. Three weeks from failure, maybe five. In most utilities today, that signal is either missed, ignored, or rides a ticket queue for a week before someone with the right context looks at it. In a multi-agent setup, this is what happens:
- A condition-monitoring agent flags the deviation against baseline and historical behavior.
- A diagnostic agent correlates with the asset’s maintenance history and narrows the likely failure to two or three components.
- An inventory agent queries the warehouse: part A is at the regional depot, part B only in central stock.
- A procurement agent checks supplier catalogs, lead times, and qualified alternatives; reserves stock pending human approval.
- A planning agent respects operational constraints - this pump cannot go offline during peak demand - and proposes a maintenance window that fits.
A human operator gets one decision to make, not twelve. And the decision arrives with the context needed to make it. The same shape of problem repeats across energy distribution, rail, manufacturing, building management — anywhere asset health meets supply chain meets a maintenance crew’s calendar. These are precisely the problems that don’t belong to any one department, which is why they stay difficult.
The economics aren’t subtle.
Unplanned downtime in water and energy infrastructure costs between five and ten times what planned maintenance costs — before you count regulatory exposure for service level misses. Working capital tied up in “just in case” spare parts, across asset-heavy industries, runs into the billions. Most of it sits there because no one trusts the forecast enough to hold less. Agents that cross the seams of the enterprise change both numbers. Not by being smarter than humans. By being more patient, more systematic, and — critically — awake at 3 a.m.
The architecture is less exotic than the hype suggests.
Each agent operates in four layers: a goal it optimizes for, a set of playbooks describing how similar problems have been handled before, a set of skills (learned procedures), and a set of tools (API calls, queries, writes). They share a runtime: retrieval for domain knowledge and SOPs, a communication layer for handoffs, an audit trail of every action, a budget engine that stops them running away with themselves, and configurable autonomy per action — from “suggest and wait” to “do it, tell me after.” Agents do not share memory. They share structured artifacts. That is the critical design choice, and it is also what makes the whole thing auditable — which is what makes it acceptable to anyone accountable for a regulated asset.
Now the unpopular part.
The limiting factor in industrial multi-agent systems is not the model. It is not the prompt. It is your API landscape. SCADA and condition monitoring: OPC UA is the standard anyone sensible is converging on. Much of the installed base still speaks Modbus, proprietary drivers, and Excel. CMMS (SAP PM, Maximo, IFS): APIs exist. Documentation is a negotiation. What you can actually write back is politically scoped. ERP inventory: APIs are fine. Semantics are the problem. “Available” means at least four different things depending on whether you ask the planner, the buyer, or the warehouse lead. Procurement and supplier catalogs (OCI, BMEcat, PunchOut): heterogeneous, rarely machine-actionable end to end. The maturity of your agents is capped by the maturity of your APIs. Not by anyone’s LLM. This is unpopular to say because it reframes an AI conversation as an integration conversation - which is less glamorous, and more expensive. It is also, empirically, where the projects I watch live or die. If you build business APIs for a living, the requirement is clear: documented, idempotent, fine-grained scopes, and above all semantic clarity. An agent can handle ambiguity. It cannot handle three different sources of truth that each call themselves authoritative.
The interesting part of multi-agent systems is not the agents.
It is what becomes possible when they can act across the seams of your enterprise — the seams that no human team has the time to properly own. Most organizations, building this, will discover that their real inventory is not parts. It is context. At EACG we are building exactly this pattern for software compliance operations on top of TrustSource — happy to trade notes with anyone running into it on the industrial side.