Cohere proposes a fix for agent sprawl: competitive selection instead of bans

The proliferation of AI agents inside companies has turned out to be a management problem, not a technical one. When any team can build and deploy its own agent in an hour, experimentation takes off — but so does chaos: dozens of similar solutions, duplicated functionality, and lost visibility into data and access. Cohere calls this state agent sprawl, and instead of banning duplication, it proposes forcing competing agents to prove their worth.

Why agent proliferation is actually normal

Decentralized agent development has real upsides. People working directly with a process see opportunities for automation better than a central team does. When a useful agent becomes reusable, a solution built by one team ends up saving others work. At the company level, a growing number of agents can be a sign of active engagement with new ideas and the search for AI applications.

An important caveat: more agents doesn't mean higher token spend. An unused agent consumes no tokens. The problem arises when agent governance falls behind the number of agents in use.

Three models for governing agents

In practice, companies choose one of three approaches:

Centralized control — a central AI or engineering team builds and manages all agents. The upside is control; the downside is that it slows innovation.

Federated approach — business units and teams build their own agents, while a central function provides infrastructure and standards. More flexibility, but harder to govern.

Hybrid governance — the center and business units share responsibility for development and operations under common standards.

Nearly all of these models rely on agent registries — logs that track ownership, usage, access rights, and help employees find existing solutions. The logic of preferring reuse over building something new is obvious: fewer agents to monitor, lower governance overhead, lower data-access risk.

But this logic hides a catch: if a working agent already exists, the company bans building alternatives — even when a team is confident it can build something cheaper, more reliable, or more specialized. That means the company loses the chance to improve.

Competitive selection instead of bans

Cohere proposes a fundamentally different path: allow duplication, but make it conditional.

If a team has solid reasons to believe it can improve on an existing agent (the incumbent), it's allowed to build a competing variant (the challenger). But the challenger's survival depends on results: does it win a direct, metrics-based comparison?

The process calls for a time-boxed, fact-based comparison. The specific metrics depend on the task, but typically include reliability, token usage, latency, and output quality.

Once enough data has accumulated, the organization makes a call: replace the incumbent with the challenger, shut down the challenger, or keep both (if each performs better on different subtasks).

Example: a bank rethinks credit memos

A bank's commercial division uses an agent that drafts credit memos for relationship managers. A team working on scheduled annual reviews builds an alternative: fewer parameters, cheaper to run, plus an extra check that verifies figures against source documents.

For 30 days, both agents run in parallel. The bank compares them on four metrics: the percentage of figures correctly traced back to source documents, how heavily credit officers edit each version, the token cost per memo, and the time to first draft.

If the challenger wins on quality while delivering significant savings, it replaces the incumbent. If it loses without offsetting benefits, it gets shut down. If it only wins on scheduled reviews, both agents stay — each in its own role.

The key to this approach: allow duplication for the sake of experimentation, but add pressure to make a selection. That way, companies preserve their pace of innovation without falling into sprawl.