The distinction matters if you already have a function that works. An order-processing function, for instance, can still reject malformed requests, check required fields, and apply fulfillment rules in Python. It can then ask a model to flag unusual details and send uncertain cases to a human reviewer. The binding takes care of setup; your code still decides what the model sees, when to call it, and whether to act on the answer.
Two bindings, two ways to organize the work
Hosted Skill binding (app.hosted_skill). Choose this when you want a skill-oriented interface rather than provider-specific agent and session APIs. Your handler can call skill.run(), stream results with skill.stream(), or continue a conversation by supplying a session_id. In the Hosted Skills model, an .agent.md file pairs YAML configuration with Markdown instructions. A SKILL.md file, despite the similar name, is reusable guidance rather than the hosted skill itself.
Agent bindings (app.markdown_agent). Choose this when the Python handler runs the workflow and needs direct access to an SDK. The binding supplies a configured agent or session as a handler parameter and manages its invocation lifecycle. Here, .agent.md is read as instructions only: the app configures the client, model, and tools separately. Microsoft’s October Azure SDK Blog described a preview that supported only Microsoft Agent Framework; the Ignite announcement also lists the GitHub Copilot SDK. That makes the older post useful background, but not the current provider list.
What to check before a pilot
- Check HTTP compatibility before reusing an app. Microsoft Learn says a hosted-skills app cannot mix streaming and non-streaming HTTP functions. That is a constraint on the hosted-skills app model, not a blanket rule for Python agent bindings. If an existing app has ordinary HTTP handlers, verify which model you’re adopting before changing those routes.
- Treat the function app as a boundary for shared tools. Microsoft Learn says every agent binding in an app receives all automatically discovered skills and Model Context Protocol (MCP) servers. The preview does not let individual bindings select a subset. If a read-only summarizer and an agent with write access need different capabilities, Microsoft recommends putting them in separate function apps.
- Keep model calls out of orchestration replay. With the optional Durable Functions integration,
context.call_agent()runs agent work in an activity rather than replaying model and tool calls in the orchestrator itself. Inputs and outputs must be JSON-serializable. That keeps orchestration replay deterministic; it does not make the model’s answer deterministic.
A useful first test is a classification or summarization step that can fail without blocking the rest of the workflow. Check what the function does when the model times out, returns an unusable result, or suggests an action the application would otherwise reject. The preview can save setup code. The better test is whether the function still behaves predictably when the AI doesn’t.
References
- (In preview) Public Preview: Hosted Skill and Agent Bindings for Azure Functions Azure Updates · 2026-11-17T20:00:32Z
- Bring agentic reasoning to Python function apps with Azure Functions Agent bindings (preview) - Azure SDK Blog devblogs.microsoft.com
- Agent bindings for Python in Azure Functions | Microsoft Learn learn.microsoft.com