Microsoft Foundry Agent Service | Robots Atlas

Description

History and rebrand

The service debuted in November 2024 as Azure AI Agent Service — the Azure AI Foundry component responsible for agent runtime. It reached public preview in March 2025 and became generally available in late 2025. In April 2026, as part of a broader brand consolidation, Microsoft renamed the platform to Microsoft Foundry Agent Service, and Azure AI Foundry itself became Microsoft Foundry. Legacy endpoints (`*.services.ai.azure.com/api/projects/.../agents`) remain backward compatible.

Platform architecture

Foundry Agent Service supports three agent types: Prompt agents (no-code, configured in the portal or via SDK), Workflow agents (declarative YAML orchestrations with branching and human-in-the-loop), and Hosted agents — code-based agents built with any framework (Microsoft Agent Framework, LangGraph, custom code) and deployed as containers on isolated Micro VMs that scale per session.

Tools and protocols

Built-in tools include web search, file search, memory, code interpreter, and hosted MCP servers (including Azure DevOps MCP). Toolbox lets teams publish a tool set once and consume it from any MCP-compatible runtime. Agents authenticate through Microsoft Entra (managed identity, OAuth On-Behalf-Of). Publishing supports OpenResponses, Activity, and A2A (preview) protocols for agent-to-agent communication, while distribution targets Microsoft 365 Copilot, Teams, and the Entra Agent Registry.

MLOps Lifecycle

7/10 supported

Prompt Management

Prompt registry — central prompt repository

Versioning — prompt versioning and history

Testing frameworks — A/B testing and prompt evaluation

3 / 3 supported · none unsupported

Monitoring

Data drift detection — input data drift detection

Concept drift detection — concept drift detection

Hallucination monitoring — LLM hallucination monitoring

Bias evaluation tools — bias evaluation tooling

3 / 4 supported · 1 unsupported hidden

Human-in-the-Loop

Labeling services — data labeling tools

RLHF workflows — reinforcement learning from human feedback

Manual override — manual override of model decisions

1 / 3 supported · 2 unsupported hidden

Data & Knowledge

Applications

11

Architecture & Mechanisms

19

Security

Developer Ecosystem

SDK Languages

PyPythonJSJavaScriptTSTypeScript

API Type

REST

Community & resources

Templates library

Quickstarts

API Reference

Tutorials

Pricing & Business Model

See full pricing

Pricing models

Usage-based

Token-based

Provisioned throughput

Resource quotas

Per project

Per user

Cost alerting

SLA & Support

99.9%uptime SLA

StandardEnterprise 24/7

Supported AI Models

17

Description

History and rebrand

Platform architecture

Tools and protocols

MLOps LifecycleiMLOps LifecycleFull model lifecycle: registry, feature store, prompt management, monitoring and human-in-the-loop.

Prompt Management

Monitoring

Human-in-the-Loop

Data & KnowledgeiData & Knowledge ManagementData connectors, vector database integration, native vector search and data management (PII, provenance, synthetic data).

ApplicationsiAI ApplicationsDomains and use cases this platform is best suited for — from RAG and fine-tuning to scientific research.

Architecture & MechanismsiArchitecture & MechanismsArchitectural foundations and modern AI processing methods that are natively supported or used by this platform.

SecurityiEnterprise SecurityCertifications, access controls and data-protection features essential for corporate deployments and cloud privacy compliance.

Developer EcosystemiDeveloper EcosystemDeveloper resources: available SDKs, supported programming languages, and infrastructure features and model-deployment methods.

Pricing & Business ModeliPricing & Business ModelBilling models (usage-based, provisioned throughput), resource limits and SLA parameters (uptime, support tiers).