ディスカバリーから実行まで：Microsoft FoundryのToolboxとRoutinesでエージェントをスケールDiscovery to Execution: Scaling Agents with Toolboxes and Routines in Microsoft Foundry

Microsoft has outlined two capabilities within Microsoft Foundry, Toolbox and Routines, that are designed to help development teams scale AI agents from early prototypes to production systems. The framing in the source is direct: tooling rarely breaks at a small scale, but it tends to break when teams move to production, where reliability, governance, and security become non-negotiable. As AI adoption accelerates, the volume of tools that agents can reach has grown in parallel, and discovering, managing, and securing the right ones is emerging as a central challenge for engineering teams.

The underlying problem is structural rather than cosmetic. An agent is only as useful as the actions it can take, and those actions are typically exposed as tools: a database query, a call to an external API, a document search, or a trigger into another business system. A handful of tools is easy to wire up by hand. A catalog spanning dozens or hundreds of tools across multiple teams is not. At that point, developers face questions about which tool is the correct one for a task, who is permitted to call it, how credentials are handled, and how usage is audited. The source suggests these issues appear most acute when teams cross from controlled demonstrations into live environments where access control and traceability carry real consequences.

Toolbox is presented as a way to address the discovery and management side of that equation. Conceptually it functions as a centralized place where tools can be registered, found, and governed, so that agents and the people building them are not left assembling integrations ad hoc. The emphasis on securing tools is notable, because exposing an agent to powerful actions without clear permission boundaries is a recognized risk in agentic systems. By centralizing tool definitions and the policies around them, the approach appears intended to reduce duplicated effort and give organizations a clearer view of what their agents can actually do.

Routines, by contrast, focus on execution. Where Toolbox concerns the inventory of available capabilities, Routines appear aimed at codifying how agents carry out multi-step work in a repeatable way. The pairing reflected in the title, from discovery to execution, captures the intended workflow: teams first locate and authorize the right tools, then compose reliable, reusable sequences that can run consistently in production. This separation of concerns is a common pattern in software platforms, and applying it to agents is a reasonable response to the unpredictability that can accompany large language model driven automation.

It helps to place these features in the broader context of Microsoft Foundry, the platform previously known as Azure AI Foundry, which Microsoft has positioned as its environment for building, evaluating, and deploying AI applications and agents. Foundry sits alongside related products such as Copilot Studio for lower-code agent creation and the various Azure model and orchestration services, so Toolbox and Routines are best understood as additions to an existing stack rather than standalone products. Their value is likely to depend on how tightly they integrate with the model catalog, evaluation tooling, and identity controls already present in that ecosystem.

The announcement also aligns with an industry-wide shift toward standardizing how agents connect to tools and data. The Model Context Protocol, an open standard for linking models to external systems, has gained traction across several vendors, and Microsoft has signaled support for interoperable tool connectivity. A centralized toolbox concept fits naturally with that direction, since a registry is more useful when the tools it catalogs speak a common interface. Competing platforms from other major cloud and AI providers have been pursuing comparable goals around tool management, agent orchestration, and governance, which suggests this is a converging area of investment rather than an isolated move.

For teams evaluating the offering, the practical takeaways center on operational maturity. The features are pitched at the moment when experimentation gives way to deployment, and their benefit will be measured by how much they reduce integration overhead, enforce security boundaries, and make agent behavior predictable. As with any newly described capability, the specifics of availability, pricing, and regional support warrant confirmation against Microsoft's official documentation, since blog descriptions tend to emphasize intent ahead of complete implementation detail.