Microsoft Releases Stable Agent Harness for Python and .NET

Microsoft has shipped a stable agent harness as part of its Agent Framework, targeting both Python and .NET developers. The harness provides runtime scaffolding that turns a language model into an agent capable of calling tools, managing multi-step tasks, retaining memory, and handling approvals—all with defaults that can be changed.

An agent harness is the layer between a raw chat model and a functioning agent. A model on its own generates text; the harness adds function invocation, planning, memory, context window management, and telemetry. Microsoft’s implementation bundles these features into a single wrapper that developers can apply to any chat client. The result is a customizable pipeline for long-running work such as research, data analysis, and general task automation.

The harness includes function invocation with a configurable iteration limit, per-service-call history persistence for crash recovery, compaction to avoid overflowing the context window during long tool loops, a persistent todo list with plan/execute mode tracking, durable file memory, progressive skill loading, optional web search when the underlying inference service supports it, tool approval with standing rules and heuristic auto-approval, and built-in OpenTelemetry telemetry. All features are enabled by default and individually removable.

Developers supply only a chat client, instructions, and optionally custom tools. In .NET, the harness is added with a single .AsHarnessAgent() call on an IChatClient. In Python, a create_harness_agent() function wraps a FoundryChatClient. Microsoft provides examples of research assistants that plan, search the web, and execute tasks; data-processing agents with approval-gated file tools; and domain-specific assistants like a personal-finance agent that combines tools, memory, skills, and planning.

While the core harness is stable, Microsoft is holding back several opt-in features it considers not yet ready for release: background agents that delegate subtasks concurrently, scoped file access tools, automatic re-invocation loops until a completion condition is met, and shell tooling. Using these features triggers a warning, and the company says it wants more customer feedback before shipping them.

The harness is closely tied to Microsoft Foundry (formerly Azure AI Studio) for model hosting, but the pattern itself is client-agnostic. Developers using other chat clients can still apply the harness, though the built-in web search and some integrations may require Foundry’s infrastructure.

Microsoft’s release provides a runtime for developers who would otherwise stitch together their own tool-calling loops, memory stores, and approval workflows. The trade-off is that the harness comes with a set of opinions about how agents should behave—opinions the company acknowledges are customizable but still baked in. Developers can evaluate whether those defaults suit their use cases and whether the withheld features will be refined based on feedback.