Agent mesh network or a single agent deployment — which does an enterprise actually need?

The answer

The question that decides it

Will this agent ever need to transact with an agent your organisation does not control? If the honest answer is no — the workflow is internal, the systems are yours, the approvals are yours — a mesh adds discovery and identity machinery your one agent does not need. Build the single deployment: an authority model, an evaluation suite, an audit trail, and a rollback, which is the whole of what most enterprise AI programmes actually require.

When the mesh earns its cost

Coordinating with a supplier’s agent on a purchase order. Settling a claim with a partner’s agent without a shared platform intermediating both of you. Any case where the other party would not accept your platform as the broker, because it is a competitor, a regulator, or simply not obligated to. That is a structural argument for a mesh, not a preference for one.

A staged answer beats a premature one

The lowest-risk path is a bounded single-agent pilot first — proving the authority model and the evidence trail on a workflow you fully control — and only then extending into mesh coordination once the internal discipline is proven. A mesh deployment that skips this step inherits an ungoverned agent with more counterparties attached to it.