Microsoft shifts focus to AI agents with 11,000-model catalog

From chat to delegation

Microsoft CEO Satya Nadella is pivoting the company's strategy toward AI agents. Instead of focusing on which specific chatbot is best, the company wants to be judged on what its software can actually do for users. The goal is to move away from simple chat windows where a user waits for a single response.

The new direction focuses on 'macro delegating' and 'micro steering.' In this model, a user sets a high-level goal and the software performs the long middle steps of a task. A human then steps back in only to provide judgment or direction when the agent reaches a specific milestone.

This represents a shift toward asynchronous work. While a chat assistant answers a prompt immediately, an agent might take a brief and return hours later with a completed result. This allows the software to handle complex, multi-step processes that would be too tedious for a human to prompt manually.

The swappable model engine

A core part of this strategy involves a catalog of more than 11,000 models. This includes models from major competitors like OpenAI, Anthropic, Mistral, and xAI, alongside Microsoft's own MAI family. By hosting nearly everyone, Microsoft aims to make the specific model behind an agent a swappable part.

Microsoft is building a system where the 'harness' stays constant while the model changes. The harness includes the context, memory, and action space—the rules and tools the AI can access. By separating these from the model family, Microsoft wants to own the plumbing of the AI experience rather than just the engine.

This approach helps enterprise buyers who need to test different models against their own data. It also allows a company to swap out a model if a provider raises prices or experiences a supply problem. As a result, the model becomes a commodity input while the orchestration layer remains the primary product.

Governance and control planes

As agents begin to act on behalf of users, Microsoft is introducing Agent 365 as a control plane. In cloud computing, a control plane sets policies while the data plane does the actual work. This registry will help security teams track which agents exist, who owns them, and what permissions they have to call tools.

This governance is necessary because agents can hold credentials and perform actions autonomously. Without a central registry, companies might struggle with 'shadow IT' where agents are deployed without oversight. A dedicated control plane allows for identity management, permission logs, and a way to audit what an agent changed during its run.

The push for governance also responds to outside pressure from regulators. Software that acts without a person clicking every step invites scrutiny from international bodies. By shipping a control layer alongside the agents, Microsoft can sell compliance as a standard part of its enterprise package.

Real-world adoption and scale

The scale of this transition is already visible in large-scale deployments. Levi Strauss & Co. reportedly used Microsoft's platform to deploy more than 1,000 domain-specific agents. These are not single general assistants but hundreds of narrow workers, each tied to a specific internal function or process.

This deployment shows how the multi-model pitch works in practice. A large retailer does not necessarily need the single smartest model for every task; it needs the right model for each specific job. By using both OpenAI and Anthropic models, the company can optimize for different needs while keeping the same infrastructure.

However, the success of these deployments depends on the reliability of the 'loop' around the model. A strong model can still fail if the instructions or memory layers are sloppy. Microsoft's bet is that by owning the orchestration layer, they can provide the reliability that multi-step tasks require.

The shift in public perception

The move toward agents aligns with a broader shift in how AI is being described in the public sphere. Recent executive orders in Washington have directed federal agencies to use the term 'Super Intelligence' instead of 'Artificial Intelligence.' While this is currently a naming rule, it reflects a growing view of AI as a system with high capabilities.

Both the political shift in language and Microsoft's technical shift toward agents suggest a move from tools to actors. When software is described as something that 'works on our behalf,' it changes the expectation of what the technology does. It moves the focus from providing answers to achieving outcomes.

As these systems become more autonomous, the focus for users and regulators will shift toward responsibility. If an agent is delegated a task and makes an error, the question of who is accountable becomes a legal and insurance topic. Microsoft's focus on a control plane suggests they are preparing for this transition from assistance to agency.