Microsoft and Notion shipped agent infrastructure in the past three weeks, from entirely different starting positions, and arrived at the same product category that Anthropic has spent eighteen months building. Frontier models still advance, but the products that these companies chose to build and sell sit between the model and the world. Harnesses, protocols, and control surfaces that determine what an agent can reach, run, and report have become the primary competitive layer.

Infrastructure (has to) come first

Claude Code reached the top position among AI coding tools within eight months of its launch, holding a 46% favorability rating in recent independent surveys. Anthropic’s engineering team attributed that trajectory to the harness, the orchestration layer that manages tool dispatch, permissions, context windows, and session state, as much as to the underlying model.

Anthropic generalized that harness into a product line. MCP, launched in November 2024, standardized tool connectivity as an open protocol. Thousands of published servers now exist, and every major AI provider ships MCP-compatible tooling. The Claude Agent SDK, announced in January, made the same architecture reusable for non-coding agents. Managed Agents, which followed in April, virtualized the full stack into three swappable interfaces: session (the append-only event log), harness (the orchestration loop), and sandbox (the execution environment). Each component can fail or be replaced independently.

Anthropic’s engineering blog framed Managed Agents as operating-system design, virtualizing abstractions general enough for programs that did not yet exist. The harness encodes assumptions about what Claude cannot do alone. Those assumptions go stale as models improve, which makes the interface more durable than any particular implementation behind it. The pattern held across surfaces. The same Agent SDK now runs Claude Code in the terminal, as extensions for VS Code and JetBrains, as a GitHub Action, and inside claude.ai itself.

Control planes and workspaces

Microsoft Agent 365 launched on May 1 at $15 per user per month as a governance and security control plane for enterprise AI agents. The product discovers, inventories, and governs agents across Microsoft platforms, AWS Bedrock, and Google Cloud. It detects unmanaged agents running on Windows devices, including Claude Code, and treats shadow AI as an asset-management problem. In June, Defender and Intune will gain the ability to block unsanctioned agent execution at the endpoint, bringing local AI agents under the same policy framework that governs human access to enterprise resources. Microsoft bundled Agent 365 into a new E7 Frontier Suite at $99 per user per month, its first enterprise licence tier since E5 launched in 2015.

Notion shipped its Developer Platform on May 13, adding Workers (a hosted runtime for custom code), an External Agents API, and database sync from any API-connected source. Workers pair deterministic custom code with LLM flexibility, letting teams enforce exact business logic without relying on model reasoning for steps that should execute the same way every time. External agents including Claude Code, Cursor, Codex, and Decagon now appear as workspace participants that can be assigned tasks and tracked alongside human collaborators. Over one million Custom Agents have been created since Notion launched the feature in February.

Microsoft approached from enterprise security, Notion from collaborative workspace, and both concluded that the model is a component. The orchestration surface determines which agents an organization can deploy, and identity, governance, tool routing, and data sync have become the differentiating features.

Born in the saddle

OpenAI released GPT-5.5 in late April with design priorities that serve harness infrastructure. The model recovers from errors mid-task, makes more efficient tool calls, maintains coherence over longer contexts, and shows better calibration, meaning that it less often proceeds confidently with a bad plan.

These improvements make GPT-5.5 a more reliable component within whatever orchestration infrastructure surrounds it. OpenAI described the model as built for tasks in which a model must plan, use tools, check its work, and keep going. Those capabilities presuppose a harness to coordinate them. A model designed to recover from tool-call failures and self-correct mid-task fits into the orchestration layers that Anthropic, Microsoft, and Notion are building.

The scaffolding is the strategy

In prior platform shifts, the valuable layer migrated from the core technology to the orchestration infrastructure once that infrastructure stabilized. The frontier model race continues, and capability still matters. The commercial contest over the next year, however, will center on who owns the scaffolding that connects those models to the systems in which enterprises actually do their work.

Subscribe now

AI Central

The Platform Shift Picks Up Steam