Cohere says enterprises need ways to manage growing numbers of AI agents as employees gain tools to build and deploy agents for their own work. Without a shared view of what exists, teams may struggle to find and reuse agents, while governing access to company data and systems becomes harder. Cohere calls proliferation “agent sprawl” when management cannot keep pace with agent creation.

Decentralized experimentation can help employees closest to a workflow build agents tailored to it, then share useful solutions with other teams. A larger agent inventory does not automatically raise inference costs: an unused agent consumes no tokens simply by existing. The burden arises when oversight falls behind. Cohere says enterprises commonly use centralized delivery, where a central AI or engineering team builds and governs agents; federated delivery, where teams own development and operation with central support for shared infrastructure and governance patterns; or a hybrid of the two.

Organizations are also using agent registries to track ownership, usage and access permissions, and help employees discover existing agents. Reuse can reduce overlapping options and the number of agents to monitor, but a blanket ban on duplication may block alternatives that are cheaper, more reliable or better suited to a task. Cohere proposes allowing a team with a credible improvement case to build a challenger agent, while making its continued use conditional on evidence of an advantage over the incumbent.

The comparison should be time-bounded and use measures suited to the work, such as reliability, token use, latency and task performance. After gathering enough evidence, the organization can replace the incumbent, retire the challenger, or keep both if each fits a different part of the workload. Cohere illustrates the approach with a bank’s commercial lending division: a central team’s finance agent drafts credit memos, while a lending team builds a challenger using a smaller, cheaper model and an extra step to check figures against source documents.

For 30 days, the bank tests both agents on cases ranging from routine annual reviews to complex financing. It compares the share of figures correctly traced to source documents, how much credit officers edit drafts, token cost per memo, and time to first draft. A challenger with comparable quality at substantially lower cost replaces the incumbent; one without a meaningful advantage is retired. If it works better for routine reviews but worse for complex financing, both remain assigned to the work each handles best. Cohere’s proposal is to make agent creation easy while making an agent’s continued use depend on the results of comparison.

Cohere says enterprises need ways to manage growing numbers of AI agents as employees gain tools to build and deploy agents for their own work. Without a shared view of what exists, teams may struggle to find and reuse agents, while governing access to company data and systems becomes harder. Cohere calls proliferation “agent sprawl” when management cannot keep pace with agent creation.

Decentralized experimentation can help employees closest to a workflow build agents tailored to it, then share useful solutions with other teams. A larger agent inventory does not automatically raise inference costs: an unused agent consumes no tokens simply by existing. The burden arises when oversight falls behind. Cohere says enterprises commonly use centralized delivery, where a central AI or engineering team builds and governs agents; federated delivery, where teams own development and operation with central support for shared infrastructure and governance patterns; or a hybrid of the two.

Organizations are also using agent registries to track ownership, usage and access permissions, and help employees discover existing agents. Reuse can reduce overlapping options and the number of agents to monitor, but a blanket ban on duplication may block alternatives that are cheaper, more reliable or better suited to a task. Cohere proposes allowing a team with a credible improvement case to build a challenger agent, while making its continued use conditional on evidence of an advantage over the incumbent.

The comparison should be time-bounded and use measures suited to the work, such as reliability, token use, latency and task performance. After gathering enough evidence, the organization can replace the incumbent, retire the challenger, or keep both if each fits a different part of the workload. Cohere illustrates the approach with a bank’s commercial lending division: a central team’s finance agent drafts credit memos, while a lending team builds a challenger using a smaller, cheaper model and an extra step to check figures against source documents.

For 30 days, the bank tests both agents on cases ranging from routine annual reviews to complex financing. It compares the share of figures correctly traced to source documents, how much credit officers edit drafts, token cost per memo, and time to first draft. A challenger with comparable quality at substantially lower cost replaces the incumbent; one without a meaningful advantage is retired. If it works better for routine reviews but worse for complex financing, both remain assigned to the work each handles best. Cohere’s proposal is to make agent creation easy while making an agent’s continued use depend on the results of comparison.