Summary: In a post published October 15, 2026, Cohere senior solutions architect Kasim Patel argues that enterprise agent sprawl is less about too many agents and more about managing them without selection pressure. The piece surveys centralized, federated, and hybrid operating models, then proposes an alternative: let challenger agents compete against incumbents on evidence, and retire the losers. For anyone shipping agent platforms, that framing changes what “governance” should mean.
Sprawl is a management failure, not a headcount problem
Cohere’s core observation is easy to miss: an unused agent costs almost nothing. As Patel’s article puts it, an agent that isn’t running doesn’t consume tokens just by existing. The overhead only appears when your process for tracking, discovering, and governing agents falls behind the rate at which people build them. That’s the moment proliferation becomes sprawl.
The management symptoms are the familiar ones. Teams rebuild agents that already exist because they can’t find them. Access to enterprise data spreads across agents nobody owns. Governance gets harder precisely because creation got easier — employees can now build task-specific agents without filing a ticket to a central AI team.
Worth reading alongside: my earlier look at Cohere’s North 2 platform, which is the product surface this argument sits on — shared, reusable agents that run inside the enterprise perimeter.
Proliferation is a signal worth protecting
Cohere is careful not to treat decentralization as the villain. The people closest to a workflow are usually best placed to see how an agent could streamline it, and when a useful agent gets shared, a fix built for one team can be adopted by another. A growing agent count can simply mean employees are actively experimenting.
That’s why the standard reflex — an agent registry plus a strong norm against duplication — only solves half the problem. Registries help with discovery, ownership, and permissions, and Cohere acknowledges enterprises increasingly rely on them. But if the existence of a workable agent blocks all alternatives, you forfeit the chance to find something cheaper, more reliable, or better fitted to the task.
Treat duplication as provisional
The alternative Cohere proposes is a challenger model. A team that believes it can beat an incumbent agent may build a variant — but the challenger’s continued existence depends on demonstrating a meaningful advantage within a time-bounded, evidence-based comparison. Metrics might include reliability, token use, latency, and task performance, depending on the use case.
The worked example is concrete. A bank’s central team ships a finance agent that drafts credit memos. A lending team builds a challenger on a smaller, cheaper model, adding a step that cross-checks figures against source documents. For 30 days both draft memos across the same range of cases. The bank compares four things: how many figures trace correctly to sources, how heavily credit officers edit each draft, token cost per memo, and time to first draft. Then three outcomes: challenger wins outright and replaces the incumbent; challenger shows no edge and gets retired; or each wins on different work — routine annual reviews versus complex financings — and both stay, each assigned to what it does best.
What this means for how you build
The generalizable principle is pairing variation with selection pressure. Make agent creation easy, but make persistence conditional. Registries and anti-duplication norms address inventory; they don’t tell you which agent should survive. A structured incumbent-challenger process does.
The honest caveat: this only works if someone actually runs the evaluations. A 30-day comparison with traceability checks and edit-rate measurement is real work, and the supplied article doesn’t specify how enterprises should operationalize the评测 infrastructure. If you already have an eval harness for your agents, that’s the natural place to hang a challenger process — start with one contested agent and see whether the evidence changes anyone’s mind.
The takeaway: sprawl isn’t solved by building less. It’s solved by making sure everything that keeps running has earned it.