Skip to main content

Microsoft Foundry + Diagrid

Build agents with Microsoft Foundry. Make their execution durable with Diagrid.

Microsoft Foundry gives your agents models, tools and a managed place to run. Diagrid gives every agent run durable execution: each model call and tool call is recorded as a step of a workflow, so a run survives crashes and restarts and resumes where it stopped.

The two are complementary. You keep your agent framework, your Foundry models, your tools and your business logic, and add one Diagrid package to the agent.

A Foundry agent with Diagrid durable executionYour agent appyour codeDiagridMicrosoft FoundryAgent frameworkLangGraph, CrewAI, …Diagrid agent packagedurable runnerModelsToolsWorkflow engineExecution historyDiagrid CLI and Consolemodel and tool callsevery step
A Foundry agent with Diagrid durable executionYour agent calls Foundry for models and tools, and records every step of the run in Diagrid. The agent app runs on your machine, in your cloud, or as a Foundry hosted agent.

What each platform provides​

Microsoft FoundryDiagrid
Models and toolsModel catalog, model deployments and toolsWorks with the models and tools your agent already uses
Agent runtimeHosted agents: a managed sandbox, endpoint and identity for your agent containerNever runs your agent code. Your agent connects to Diagrid over outbound HTTPS and gRPC
ExecutionRuns the model and hosts the agent processRuns the agent's control loop as a durable workflow and stores its history
VisibilityTraces and logs of the hosted agentEvery run as a workflow: status, steps, inputs, outputs and timing

What Diagrid adds to a Foundry agent​

  • Durable execution with automatic failure recovery. Each model call and tool call runs as a workflow activity. A failed activity is retried, and when the agent process stops, the run resumes from its last completed step after a restart.
  • Persistent execution history and checkpoints. Diagrid stores the result of every completed step in the project's workflow state store. A resumed run reads those results from the history and doesn't call the model or the tool again for them. Activities run at least once, so keep tools with side effects idempotent.
  • Long-running agent execution. A run isn't tied to one process or one request. It can outlive the instance that started it, which matters for platforms that scale an idle agent to zero.
  • Workflow-level visibility. Every agent run is a workflow that you can inspect with the Diagrid CLI or in the Diagrid Console, with the input and output of each step.
  • Multi-instance coordination. Several instances of the same agent can connect with one identity. Diagrid distributes the steps among them and keeps one history per run.
  • Verifiable execution. In a project with workflow history signing, Diagrid signs every history event, which gives you a tamper-evident record of what ran, and attests results that cross identities. You turn it on when you create the project. See Verifiable execution.

Supported frameworks​

The quickstart has a tab for each of these frameworks. They are Python frameworks unless the table says otherwise. Every combination was run against a real Foundry model deployment and a real Diagrid project.

FrameworkDiagrid packageTested withFoundry model access
LangGraphdiagrid[langgraph] 0.5.3langgraph 1.2.22, langchain-openai 1.8.0Microsoft Entra ID
Microsoft Agent Framework (.NET)Diagrid.AI.Microsoft.AgentFramework 1.2.1Microsoft.Agents.AI.Foundry 1.5.0Microsoft Entra ID
Deep Agentsdiagrid[deepagents] 0.5.3deepagents 0.7.38Microsoft Entra ID
OpenAI Agents SDKdiagrid[openai_agents] 0.5.3openai-agents 0.20.0API key
CrewAIdiagrid[crewai] 0.5.3crewai 1.15.41API key
Google ADKdiagrid[adk] 0.5.3google-adk 1.39.1API key
Pydantic AIdiagrid[pydantic_ai] 0.5.3pydantic-ai-slim 2.31.1API key
Strandsdiagrid[strands] 0.5.3strands-agents 1.65.0API key
Mastra (TypeScript)@diagrid/agent-mastra 0.2.0@mastra/core 1.83.0API key
Dapr Agentsdapr-agents 1.1.1Diagrid conversation componentAPI key, held by Diagrid

The Claude Agent SDK only runs Claude models, so it needs a Claude deployment in Foundry. That combination isn't verified yet and has no tab in the quickstart.

Next steps​