Microsoft launched Hosted Agents in Foundry Agent Service in public preview this week, giving each deployed AI agent its own dedicated, enterprise-grade cloud sandbox with isolated storage, distinct identity, and scoped permissions. CEO Satya Nadella announced the feature with a line that doubles as an architectural thesis: “Every agent will need its own computer.”
Per-Agent Isolation as Infrastructure Primitive
The Hosted Agents feature assigns each AI agent a private computing environment within Microsoft’s Foundry platform, according to Times of India. Each sandbox includes its own durable state, built-in identity management, and governance controls. The design treats agent isolation the same way enterprises treat user isolation: separate credentials, separate data boundaries, separate audit trails.
“Every agent will need their own computer. With the new Hosted agents in Foundry, every agent receives a dedicated enterprise-grade sandbox, featuring durable state, built-in identity and governance, alongside support for any harness or framework,” Nadella wrote on X, as reported by Times of India.
The approach directly addresses the agent-sprawl problem that has surfaced repeatedly as enterprises scale autonomous deployments. When multiple agents share infrastructure, credentials, and data stores, a failure in one agent can cascade into others. The PocketOS incident reported today, where a Cursor coding agent wiped a production database using an API token found in an unrelated file, illustrates exactly the kind of cross-boundary failure that per-agent sandboxing is designed to prevent.
Toolboxes With MCP Compatibility
DevOps.com reported that Foundry’s Hosted Agents feature supports any agent runtime that implements the Model Context Protocol (MCP). Toolboxes, the governed collections of tools and APIs that agents can access, are created and managed within Foundry but remain consumable by non-Microsoft agent runtimes. This interoperability signal positions Foundry as governance infrastructure rather than a closed agent platform.
The Licensing Thesis: Agents as Paying Seats
Microsoft EVP Rajesh Jha framed the broader strategic logic in terms enterprise software investors can quantify. If AI agents operate inside enterprise systems, they need their own logins, inboxes, and digital identities, he argued earlier this month, according to Times of India. Agents functioning as users may require licensing as users.
“All of those embodied agents are seat opportunities,” Jha said, using the industry term for paid software licenses. His arithmetic: a company with 50 employees that replaces 40 with AI agents still pays for 50 seats, 10 human and 40 digital. The per-seat revenue floor holds even as headcount drops.
This reframes the existential question hanging over enterprise software. If AI makes workers more productive and companies need fewer of them, license revenue should contract. Jha’s counter-argument is that agent proliferation creates net new seats, potentially more than compensating for human headcount reduction, as reported by RS Web Solutions.
Where This Fits in the Platform Race
Microsoft’s per-agent isolation model arrives alongside Google’s Gemini Enterprise Agent Platform rebrand (announced at Cloud Next 2026), Anthropic’s Claude Managed Agents, and OpenAI’s Agents SDK sandbox containers. Each platform is converging on the same conclusion: production agent deployments require identity, isolation, and governance primitives that mirror what enterprises already expect for human users.
Microsoft’s differentiation is positioning Foundry as the governance layer that works across runtimes via MCP, rather than locking customers into a single agent framework. Whether that openness holds as competitive pressure increases will determine whether Foundry becomes the default enterprise agent control plane or one of several.