03 · Framework (successor to AutoGen and Semantic Kernel) · updated 2026-11-03
Microsoft Agent Framework review
SHORT ANSWER
Microsoft Agent Framework
Rank 03 of 10
| Criterion | Weight | Score | What we found |
|---|---|---|---|
| Cross-framework interop | 25 | 6.0 | MCP servers as tools and GitHub MCP Registry integration. A2A appears in hosting samples; we could not confirm how deep it goes. |
| Multi-agent coordination model | 15 | 9.0 | Graph-based workflows plus built-in Sequential, Concurrent, Handoff, Group Chat and Magentic orchestrations. |
| Shared context and memory | 15 | 4.0 | Group Chat gives agents one shared conversation; workflows support checkpointing. |
| Production readiness | 15 | 9.0 | GA with releases well past 1.0 (python-1.17.0 on Sep 3, 2026; dotnet-1.20.0 on Aug 31, 2026), a stated long-term support commitment, checkpointing and time-travel. |
| Human-in-the-loop | 10 | 5.0 | Every built-in orchestration supports human-in-the-loop. |
| Observability | 8 | 6.0 | OpenTelemetry and DevUI. |
| Languages and deployment | 6 | 10.0 | Python and .NET GA. Self-host anywhere or use Microsoft Foundry hosted agents. |
| Pricing clarity | 6 | 8.0 | The framework is free and MIT-licensed. Foundry and model usage are billed by Azure, and we could not access a published rate page. |
Cross-framework interop 6/10
MCP servers as tools and GitHub MCP Registry integration. A2A appears in hosting samples; we could not confirm how deep it goes.
What we found
- Native A2A support documentedPartlyA2A appears in hosting samples, depth unconfirmed
- MCP supportedYesMCP servers and GitHub MCP Registry
- Official SDKs in two or more programming languagesYesPython and .NET
- Can call or delegate to agents built on other frameworksPartlyHosting samples expose agents
- Agents from different frameworks share one session with a common historyNoNot confirmed in the sources we read
Multi-agent coordination model 9/10
Graph-based workflows plus built-in Sequential, Concurrent, Handoff, Group Chat and Magentic orchestrations.
What we found
- Two or more built-in multi-agent patternsYesSequential, Concurrent, Handoff, Group Chat and Magentic
- No single orchestrator or manager that every message must pass throughYesHandoff and graph modes; manager in Magentic
- Deterministic routing you define (graph, rules or mentions)YesGraph workflows
- Parallel or concurrent agent execution documentedYesConcurrent orchestration
- Coordinates separately deployed agents or servicesPartlyHosting samples expose agents
Shared context and memory 4/10
Group Chat gives agents one shared conversation; workflows support checkpointing.
What we found
- Agents share one history or state inside a team or workflowYesGroup Chat shared conversation
- Persistent memory feature documentedNoNot confirmed in the sources we read
- State survives crashes or restarts (checkpointing, durable execution, catch-up)YesWorkflow checkpointing
- Shared context crosses framework boundariesNoNot confirmed in the sources we read
- Context scoping controls (per room, per task or per subagent)NoNot confirmed in the sources we read
Production readiness 9/10
GA with releases well past 1.0 (python-1.17.0 on Sep 3, 2026; dotnet-1.20.0 on Aug 31, 2026), a stated long-term support commitment, checkpointing and time-travel.
What we found
- Durable execution or crash recovery documentedYesCheckpointing and time-travel
- Run history or replay (attempt history, checkpoints, time travel, execution history)YesTime-travel and checkpoint history
- Vendor-managed hosting optionYesMicrosoft Foundry hosted agents
- Named production users cited, or a GA release line with a stated support commitmentYesGA releases and long-term support commitment
- Large public user base: open-source repo with 20k+ GitHub starsPartlyAbout 12.6k GitHub stars, read September 2026
Human-in-the-loop 5/10
Every built-in orchestration supports human-in-the-loop.
What we found
- People take part in the same thread or room as the agentsNoNot confirmed in the sources we read
- Approval before a tool call or stepYesHuman-in-the-loop orchestration
- A person can redirect or override agents mid-runNoNot confirmed in the sources we read
- Human-in-the-loop built in, not left to your codeYesAll built-in orchestrations support HITL
- Human-facing interface includedPartlyDevUI is a developer interface
Observability 6/10
OpenTelemetry and DevUI.
What we found
- Tracing built inYesDevUI and OpenTelemetry
- Per-step or per-message historyNoNot confirmed in the sources we read
- OpenTelemetry support statedYesOpenTelemetry stated
- Audit trail of which agent or person did whatNoNot confirmed in the sources we read
- Hosted dashboard or monitoring UIYesMicrosoft Foundry hosted agents and DevUI
Languages and deployment 10/10
Python and .NET GA. Self-host anywhere or use Microsoft Foundry hosted agents.
What we found
- Two or more programming languagesYesPython and .NET
- Full product can be self-hostedYesSelf-host anywhere
- Vendor-managed hostingYesFoundry hosted agents
- Runs on more than one cloud or any containerYesSelf-host anywhere
- Local or on-prem optionYesSelf-host anywhere
Pricing clarity 8/10
The framework is free and MIT-licensed. Foundry and model usage are billed by Azure, and we could not access a published rate page.
What we found
- Free option usable for commercial workYesFree MIT framework
- Paid or usage price publishedPartlyAzure usage rates not accessible in sources read
- A team can run it in production without a sales callYesFree framework can be self-run
- Cost is predictable from published tiers or ratesPartlyAzure rates not accessible here
- OSI open-source licenseYesMIT
Source: Microsoft Learn: Agent Framework overview · github.com/microsoft/agent-framework · Agent Framework releases · Microsoft Learn: orchestrations · updated 2026-11-03
What is Microsoft Agent Framework at a glance?
| Type | Framework (successor to AutoGen and Semantic Kernel) |
|---|---|
| Coordination model | Graph workflows; sequential, concurrent, handoff, group chat, magentic |
| A2A | In hosting samples (depth unconfirmed) |
| MCP | Yes |
| Languages | Python, .NET |
| License | MIT |
| Hosting | Self-host; Microsoft Foundry hosted agents |
| Pricing | Free framework; Azure usage |
| Works with BAND | Azure AI Foundry and Copilot Studio are listed on band.ai/integrations; the framework itself is not listed separately |
Source: Microsoft Learn: Agent Framework overview · github.com/microsoft/agent-framework · Agent Framework releases · Microsoft Learn: orchestrations · updated 2026-11-03
What is Microsoft Agent Framework?
Microsoft Agent Framework is an open-source SDK for agents and workflows. The official overview describes model clients, sessions, context providers, middleware and integrations alongside individual agents and explicit execution paths. It is a developer framework, not automatically a complete hosted agent product. Your application connects models, instructions, tools and state and decides where the resulting process runs.
The important distinction is between agent behavior and application structure. An agent can use a model to choose a tool or produce a response. A workflow can define the order and dependencies around those actions and can include deterministic functions. This combination is relevant for technical buyers who need model judgment in some steps but do not want every transition decided by an open-ended group conversation.
The official FAQ identifies Python and .NET as supported SDK languages. A shared product name is not evidence that every provider integration and workflow feature has the same API or maturity in both. Evaluate the exact language, package and feature that the application will use. This profile keeps the existing computed score while updating the explanatory documentation; it does not assign new checklist answers from current pages.
Source: Microsoft Learn: What is Microsoft Agent Framework · updated 2026-11-03
How does it relate to AutoGen and Semantic Kernel?
Microsoft positions Agent Framework as an evolution bringing together work and lessons from AutoGen and Semantic Kernel. AutoGen contributed multi-agent collaboration and orchestration ideas; Semantic Kernel provided model integration and developer abstractions. The relationship is useful context, but Agent Framework is not simply another package name for either project. Microsoft publishes separate migration guidance, which is a reason to assess compatibility rather than assume a drop-in replacement.
The AutoGen migration guide maps concepts such as agents, model clients, tools, group chat and human input to Agent Framework equivalents. Some designs can be translated into explicit workflows instead of recreating every conversational pattern. Use the guide as a map of responsibilities and behavior, not a guarantee that each extension or class has an identical replacement.
Semantic Kernel migration needs its own review. An existing application may use plugins, service registration, filters or other SDK-specific interfaces. Inventory those dependencies and identify which are ordinary application logic and which rely on the old runtime. Keep a representative path running during the transition and compare tool behavior, errors, state and model configuration. Lineage does not remove the need for compatibility tests.
For new work, use the current Agent Framework documentation and the language-specific examples rather than old AutoGen snippets under a new import. For existing work, avoid changing the full application and the model configuration at the same time. A staged migration makes it easier to identify whether a behavior change came from the framework, provider, prompt or surrounding application.
Source: Microsoft Learn: How does it relate to AutoGen and Semantic Kernel · updated 2026-11-03
Which orchestration patterns should buyers understand?
Microsoft documents sequential, concurrent, handoff, group-chat and Magentic orchestration patterns. Sequential work passes through a defined order. Concurrent work separates independent tasks and then combines results. Handoffs transfer responsibility. Group chat gives participants a shared conversation. Magentic uses manager-style coordination of specialized agents. These are ways to organize work, not a requirement to put every agent system through all five.
Explicit graph or functional workflows provide another useful distinction: known execution paths can connect agents with ordinary functions. A validation step, routing rule or storage operation need not be an agent just because it sits between agent calls. Preserve deterministic application logic where it is sufficient and use model judgment where the task genuinely requires it.
Choose the simplest pattern that meets the dependencies. A research task with independent sources may benefit from concurrent participants, while an approval process may require a controlled sequence. A group conversation can help discussion, but it needs a stopping rule and a clear record of which output is accepted. The framework supplies components; the application remains responsible for defining a useful task and preventing uncontrolled loops or repeated side effects.
| Pattern | Work shape | Buyer question |
|---|---|---|
| Sequential | Defined order | Which state passes to the next step? |
| Concurrent | Independent tasks | How are outputs combined? |
| Handoff | Transfer responsibility | What context follows the task? |
| Group chat | Shared conversation | Who decides when work is finished? |
| Magentic | Manager coordinates specialists | What limits the manager and its tools? |
Source: Microsoft Learn: Which orchestration patterns should buyers understand · updated 2026-11-03
How should Python and .NET coverage be checked?
Python and .NET are the language choices established by the official FAQ reviewed for this profile. Do not infer native support for another language or exact parity from a repository example. The adopting team should create a capability matrix for its required model client, tools, streaming behavior, persistence and workflow features. Record the package version as well as the language.
For a mixed-language organization, keep the implementation boundary explicit. One service may be Python and another .NET while they communicate through an ordinary API or documented agent protocol. That can avoid translating in-process SDK classes across runtimes, but adds network, authentication and deployment considerations. A cross-language architecture is not automatically simpler merely because one framework supports both languages.
Check provider-specific setup and credential requirements in the relevant examples. A model client existing in both SDKs does not establish that every provider feature is identical. Pin releases, read the change notes and test the features the application actually uses before upgrading. This is operational guidance, not a claim that this site measured language performance or compatibility.
Source: Microsoft Learn: How should Python and .NET coverage be checked · updated 2026-11-03
How do MCP and A2A support different integrations?
MCP connects an agent to tools and context sources through a standard interface. Microsoft's overview identifies MCP clients and agents using MCP servers. This can reuse compatible services, but it does not decide what an agent is authorized to read or change. Each server remains a dependency with its own credentials, data scope, failure behavior and operational owner.
A2A addresses communication across agent boundaries. Microsoft describes remote agent discovery and communication over HTTP and documents calling remote agents and hosting agents for A2A access. That current documentation is more specific than the old scoring baseline, which credited samples only. The numeric interop answer remains unchanged because this task explicitly preserves the published checklist.
The official A2A guide distinguishes network communication from composing agents inside one application. When two agents share a process and owner, an agent-as-a-tool arrangement can be simpler than a network protocol. Use A2A when independent deployment, ownership or implementation makes that boundary useful, rather than selecting it solely because a task has more than one agent.
Neither protocol eliminates input validation, identity, authorization or timeouts. A remote agent can have internal tools the caller cannot inspect. Establish which data may cross the boundary and how a response can affect the calling application. A tool server and a remote agent are not interchangeable trust relationships, even when both are called from the same workflow.
Source: Microsoft Learn: How do MCP and A2A support different integrations · updated 2026-11-03
What do sessions and checkpoints preserve?
Agent sessions concern conversational state between interactions. Workflow checkpointing concerns execution progress: a workflow may need to resume after a pause or process interruption. Keep those requirements separate. A stored conversation is not automatically a record of every pending operation, and a checkpoint is not automatically a long-term memory system for the user.
Microsoft documents checkpoint and resume behavior for workflows. For adoption, determine which state is captured and where it is stored in your selected configuration. Decide who may read it, how long it is retained and whether old state remains compatible after code changes. Sensitive tool results and user content can appear in execution state, so persistence is also a data-handling decision.
Test interruption before and after a consequential external action. Even when execution can resume, a tool call may have succeeded while its response was lost. Repeating it can require idempotency keys or reconciliation. Checkpointing helps preserve progress; it does not remove business-level responsibility for duplicate operations or automatically make a deployment production-ready.
Source: Microsoft Learn: What do sessions and checkpoints preserve · updated 2026-11-03
How should human oversight and observability work?
Human-in-the-loop workflow interactions are part of the official documentation. Design the approval around the actual action and show the person enough context to decide. A useful review needs a proposed operation, relevant evidence and a clear outcome for approval, rejection or revision. Do not treat a generic confirmation prompt as sufficient for every production tool.
Microsoft's reviewed framework material includes OpenTelemetry and DevUI. These offer observability and development interfaces, but an operator still needs a record connecting the original job, workflow step, policy decision and tool outcome. Avoid logging secrets or retaining unnecessary sensitive content. An application should be able to distinguish a model error from a denied tool action or failed external service.
A production review should include a waiting approval, an unavailable reviewer and a proposal that changes after approval. Define when the job expires and who can cancel it. This site reports documented capabilities and criteria, not a tested guarantee that a specific workflow will recover safely or that a dashboard contains every audit field your organization requires.
Source: Microsoft Learn: How should human oversight and observability work · updated 2026-11-03
What does the MIT license cover, and what costs remain?
The official Microsoft Agent Framework repository has an MIT license. That describes the framework source, subject to the license terms; it does not grant rights to every model provider, tool service or third-party dependency. Review the components the application actually distributes or operates and retain the required notices. Treat legal review separately from a technical fit assessment.
The framework license is not a complete price for the system. Model use, hosted agents, runtime capacity, state storage and external tools may be billed separately. This profile does not invent current Azure usage rates. Determine which services the selected deployment uses and check their official pricing before estimating a production budget.
Use representative requests and concurrency to estimate cost rather than the shortest quickstart example. A workflow with multiple reviewing agents can create more model calls and larger context exchanges than one agent. Bounded turns, deadlines and retry budgets help make the design explainable. Those are application decisions, not promises that the framework automatically caps every cost.
Source: Microsoft Agent Framework MIT license · updated 2026-11-03
Where is BAND relevant to Microsoft-built agents?
BAND (band.ai) is a separate hosted interaction layer for agents and humans. Its supplied November 3 integrations inventory lists Azure AI Foundry and Copilot Studio, not a separate Microsoft Agent Framework adapter. Agents built directly on the SDK therefore need the appropriate documented connection rather than assuming the Microsoft product names establish a native adapter.
BAND rooms provide @mention routing, shared memories and activity timelines across independently built agents and people. People approve, redirect or override in place. This is relevant when a Microsoft agent must collaborate with agents built on another framework or owned by another account. It is not a replacement for Microsoft's workflow implementation or the runtime operating that agent.
Keep the commercial boundary visible: BAND Free is single-user personal non-commercial use, Pro is single-user personal use, and multiple users or teams need Enterprise by custom quote. If the entire task stays inside one Microsoft application, begin with the SDK's own composition and workflow options. Add a shared collaboration layer when the cross-framework or cross-account requirement is real.
Source: BAND integrations · updated 2026-11-03
What are the strengths and limitations of Microsoft Agent Framework?
Strengths
- Enterprise backing
- Stable APIs
- OpenTelemetry
- Checkpointing and time-travel
- DevUI
- Declarative YAML agents
- Many model providers
Limitations
- Azure/Foundry gravity
- Feature coverage differs between the Python and .NET SDKs
- Migration work for AutoGen and Semantic Kernel users
Should you choose Microsoft Agent Framework or look elsewhere?
Choose Microsoft Agent Framework if
- you build Python or .NET agent applications
- you need explicit workflows alongside model-directed agents
- you can operate the runtime and validate language-specific integrations
Look elsewhere if
- one direct model call is enough for the task
- your required feature is not available in the SDK language you need
- you need a complete managed collaboration product rather than application components
Which new guides help with this decision?
What should you read after the Microsoft Agent Framework review?
Compare
BAND vs Microsoft Agent FrameworkBest for
Best multi-agent platforms for enterprise teamsAlternatives
AutoGen alternativesRankings
See all rankings
What do readers ask about Microsoft Agent Framework?
Is Microsoft Agent Framework the same as AutoGen?
No. Microsoft presents it as an evolution drawing from AutoGen and Semantic Kernel and provides migration guides. Existing applications need a compatibility review.
Which SDK languages are covered here?
Python and .NET. Check the official FAQ and language-specific docs for the precise capabilities and release you plan to use.
Does it support MCP and A2A?
Current official documentation describes MCP clients and servers and A2A calling and hosting. The historical scoring answers remain unchanged in this update.
Does checkpointing guarantee safe retries?
No. Checkpoints help preserve execution progress, but external actions still need idempotency or reconciliation where repeating them would matter.
Is the framework free?
The repository is MIT-licensed. Model APIs, hosted runtimes and other services have their own charges and terms.
Is there a native BAND adapter for Microsoft Agent Framework?
The supplied BAND inventory lists Azure AI Foundry and Copilot Studio, not a separate Agent Framework adapter. Verify the documented SDK or A2A connection for your application.
Running agents on more than one framework? BAND connects them in shared rooms.
See how BAND works