Contents
- Fast answer
- Quick facts
- What this covers
- Why it matters
- Where to go next
- Outcomes
- Step by step
- Pick neither for new work, and pick by codebase for existing
- Token predictability is what actually separates them
- Where both of these are the wrong answer
- How to check this yourself, and what we did not test
- At a glance
- Mistakes to avoid
- FAQ
Fast answer
Both are free Microsoft frameworks and both point at the same successor, which makes this an unusual comparison. AutoGen is formally in maintenance mode: its repository states it will receive no new features or enhancements and is community managed, with only bug fixes, security patches and documentation improvements, and its README tells new users to start with Microsoft Agent Framework, which it calls the enterprise-ready successor to AutoGen. The repository showed 558 open issues when we checked on 20 August 2026. Semantic Kernel is not formally in maintenance mode, but Microsoft positions Agent Framework as its successor too, and the Semantic Kernel introduction page on Microsoft Learn still carries a July 2023 authoring date and a June 2024 update stamp with no mention of Agent Framework at all. Neither is the right place to start a new project in 2026, and both are perfectly reasonable places to have an existing one.
Facts last verified 2026-08
Quick facts
- Both frameworks
- $0, free open source
- AutoGen status
- Maintenance mode, community managed
- AutoGen open issues
- 558 shown on the repository when checked
- AutoGen guidance
- README tells new users to start with Microsoft Agent Framework
- Semantic Kernel languages
- C#, Python or Java per Microsoft's own overview
- Semantic Kernel docs date
- Introduction page authored 2023-07-11, updated 2024-06-24
- Shared destination
- Microsoft Agent Framework, also free
On this page
What this page covers
A comparison visitor should understand the tradeoff, the best-fit scenario, and the next diagnostic tool to confirm the choice.
- 01Quick verdict
- 02Pricing
- 03Project status
- 04Orchestration style
- 05Decision table
- 06Where both are wrong
Why does this matter now?
The useful output of this comparison is not a winner, it is a decision about timing. If you have a working system on either, maintenance mode means bug fixes and security patches continue and an urgent migration is cost without benefit. If you are starting something new, the vendor has told you in writing where to go, and choosing either of these instead means paying for a migration later at a moment you will not control.
Internal path
Where to go next from this page
These links are part of the A8gent learning and conversion path. Use them to move from concept, to diagnosis, to workflow build, to course.
What you should be able to do after this
- Understand the status of each project
- Choose correctly for an existing codebase
- Avoid starting new work on either
- Estimate the migration you will owe
How do you do it, step by step?
1. Read both projects' own status statements
AutoGen's repository states maintenance mode, community management, no new features, and directs new users to Microsoft Agent Framework as the enterprise-ready successor with a migration guide. Semantic Kernel's Learn introduction page carries a 2023 authoring date and a 2024 update stamp and does not mention Agent Framework. Read both before reading any tutorial.
2. Match the framework to the shape you already have
AutoGen was built around conversational multi-agent group chat. Semantic Kernel is middleware for embedding AI into an existing application, with plugins wrapping your own code. If you already have one of those shapes in production, that is the one to keep, and the comparison is about your codebase rather than about the projects.
3. Budget the token difference, which is real
AutoGen's conversational multi-agent runs consume several times more tokens per outcome than a single agent, and turn counts are emergent rather than specified, so two runs of the same task can differ several-fold. Semantic Kernel's application-embedding model does not have that failure class by default. If cost predictability matters, that is a genuine separation.
4. Estimate the migration you owe either way
Both roads lead to Microsoft Agent Framework, and in both cases it is a rewrite of the orchestration layer rather than a package rename, because the successor uses typed workflows rather than group chat and takes functions directly where Semantic Kernel needs attributes and wrappers. Budget engineering weeks and schedule it alongside work you were doing anyway.
Pick neither for new work, and pick by codebase for existing
For a greenfield project in 2026 the answer is neither, and it comes from Microsoft rather than from us. The AutoGen repository states that the project is in maintenance mode, will receive no new features or enhancements, and is community managed, with only bug fixes, security patches and documentation improvements planned. Its README directs new users to Microsoft Agent Framework, describing it as the enterprise-ready successor to AutoGen, and provides a migration guide.
Semantic Kernel is not formally in maintenance mode and continues to ship releases, but Agent Framework is positioned as its successor too. Meanwhile the Semantic Kernel introduction page on Microsoft Learn carries an authoring date of 11 July 2023 and an update stamp of 24 June 2024, and does not mention Agent Framework anywhere. That is the first page a new developer lands on, describing a position the vendor's newer material has moved past.
For an existing codebase, pick by what you already have rather than by comparing the projects. AutoGen was built around conversational multi-agent group chat. Semantic Kernel is middleware for embedding AI into an existing application, with plugins wrapping code you already own. Those are different shapes, and if one of them is already in production the right answer is to keep it and plan the move deliberately rather than to migrate sideways.
Both are free, which removes cost from the decision entirely. Semantic Kernel is MIT licensed with no fee, seats or execution meter, supporting C#, Python or Java per Microsoft's own overview. AutoGen is free open-source software on the same terms. Microsoft Agent Framework, the destination for both, is also free. So this is a question about engineering time and direction, and nothing about it turns on a subscription.
Token predictability is what actually separates them
The one dimension on which these two genuinely differ in operating cost is how predictable the model bill is, and AutoGen is much worse. Conversational multi-agent runs consume several times more tokens per completed outcome than a single agent, because the agents talk to each other and each turn resends context. That multiplier alone would be manageable if it were constant, and it is not, because conversation length is emergent rather than specified.
The consequence is that two runs of the same task can differ several-fold in consumption, and nothing in the configuration guarantees convergence. The characteristic failure is agents conversing at length without reaching a conclusion, which is expensive and produces nothing. That is the normal risk of the pattern rather than a rare pathology, and it is why budgeting AutoGen as a single figure per run is unsound.
Semantic Kernel does not have that failure class by default, because it is middleware over calls your application makes rather than a conversation among agents. Your token consumption is bounded by the calls you write, which makes it forecastable in the ordinary way: count the calls, count the tokens, multiply by your provider's published rate. Nothing flows to Microsoft's framework layer in either case, so the provider rates are the whole variable cost.
If you are running AutoGen and staying for now, the mitigations are specific and worth implementing regardless of migration plans. Set hard maximum turn limits in code. Add a termination condition tied to a concrete output artefact rather than to agents agreeing with each other. Log token consumption per run from day one so runaway cases surface in testing rather than on an invoice. And sandbox anything that executes code, because agent code execution is a genuine security exposure needing isolation rather than good intentions.
Where both of these are the wrong answer
Both are the wrong answer for new work, and the alternatives depend on what attracted you. If it was the Microsoft ecosystem, Agent Framework is the sanctioned destination, is free, and has a published migration guide. Budget that migration as engineering weeks rather than a dependency bump in either direction, because the successor replaced group chat with typed workflows and takes functions directly where Semantic Kernel needs an attribute, a plugin wrapper, a kernel and a handoff.
If it was the role-based multi-agent design specifically, CrewAI is an MIT-licensed open-source Python library with no licence, seat or execution fee, actively developed, and organised around agents with defined roles collaborating on a task. The same token warning applies with full force: multi-agent designs multiply model calls per outcome, so budget per completed outcome rather than per model call, and cap the crew size deliberately.
If what you actually want is explicit control flow rather than emergent conversation, LangGraph removes the non-converging failure class by construction, because the graph decides what happens next rather than the agents negotiating it. The library is open source and self-hosts free, with LangSmith at $0 a seat on Developer with a 1 seat maximum and 5,000 base traces or $39 a seat a month on Plus with 10,000.
And it is worth asking whether the problem needed multiple agents at all, because a surprising number of AutoGen projects exist because the pattern was interesting rather than because the work decomposed. A single well-prompted agent with good tools is cheaper, more predictable and far easier to debug. If a single agent handles your task, the OpenAI Agents SDK is free and minimal, or Pydantic AI is free and typed, and neither requires choosing a multi-agent framework at all.
How to check this yourself, and what we did not test
AutoGen's maintenance-mode notice, successor guidance and the 558 open issue count came from github.com/microsoft/autogen, and Semantic Kernel's description and documentation dates from learn.microsoft.com/en-us/semantic-kernel/overview, both read on 20 August 2026. Both frameworks are free, so nothing in this comparison turns on a price.
We did not benchmark either framework, measure token consumption on a real multi-agent run, or verify the current state of Semantic Kernel's C#, Python and Java implementations against each other. The claim that AutoGen's token consumption is unpredictable follows from its documented design, where conversation length is emergent rather than specified, not from a measurement we took.
The check that decides this is reading both projects' own status statements before any tutorial. AutoGen's README tells new users to start with Microsoft Agent Framework. Semantic Kernel's Learn introduction page carries a July 2023 authoring date and a June 2024 update stamp and does not mention Agent Framework at all. Those two pages, read in that order, answer the question faster than any comparison.
At a glance
Semantic Kernel vs Microsoft AutoGen decision table, read from Microsoft Learn and the AutoGen repository on 20 August 2026
| Decision factor | Semantic Kernel | Microsoft AutoGen |
|---|---|---|
| Licence cost | $0, MIT licensed | $0, free open source |
| Project status | Active, but a successor is positioned | Maintenance mode, community managed |
| New features | Not stated as halted | Repository states none will be added |
| Ongoing changes | Releases continue | Bug fixes, security patches, documentation only |
| Open issues | Not counted here | 558 shown on the repository when checked |
| Vendor guidance for new users | Agent Framework is positioned as the successor | README directs new users to Microsoft Agent Framework |
| Documentation currency | Introduction page authored 2023-07-11, updated 2024-06-24 | Repository notice is current |
| Design shape | Middleware embedding AI into an existing application via plugins | Conversational multi-agent group chat |
| Token predictability | Bounded by your application's calls | Emergent; turn counts are not specified in advance |
| Languages | C#, Python or Java | Python |
| Migration target | Microsoft Agent Framework | Microsoft Agent Framework, with a published migration guide |
| Right for new projects | No; Microsoft points elsewhere | No; the README says so explicitly |
What mistakes should you avoid?
- Starting a new project on either when Microsoft points new users at Agent Framework
- Migrating a working deployment urgently when maintenance mode does not break it
- Letting AutoGen agents converse without a hard turn cap or a concrete termination condition
- Assuming migration to Agent Framework is a dependency bump rather than an orchestration rewrite
- Trusting the Semantic Kernel Learn introduction page as current when it is dated 2023
- Running agent-generated code without a sandbox on AutoGen
FAQ
- Which should I choose, Semantic Kernel or AutoGen?
- For a new project, neither. Microsoft names Agent Framework the enterprise-ready successor to AutoGen in AutoGen's own repository, and positions it as the destination for Semantic Kernel too. For an existing codebase, keep whichever you already have and plan the migration for when you are making major changes anyway.
- Is AutoGen still maintained?
- In maintenance mode. The repository states it will receive no new features or enhancements and is community managed, with bug fixes, security patches and documentation improvements continuing. It showed 558 open issues when we checked on 20 August 2026. Existing deployments keep working.
- Do either of them cost money?
- No. Both are free open-source software with no platform fee, seats or execution meter. Your costs are model tokens billed by your provider and hosting. Microsoft Agent Framework, the successor to both, is also free, so this is entirely a question about engineering time and direction.
- Which has more predictable costs?
- Semantic Kernel, because AutoGen's conversational multi-agent design makes token consumption emergent rather than specified. Agents decide how many turns a task takes, so two runs of the same task can differ several-fold, and agents can converse at length without converging. Cap turns explicitly if you run AutoGen at all.
Sources & further reading
Done for you
Want this built for your business?
Tell us the workflow. We scope, build, and ship the agent with guardrails and the numbers to prove it worked. The scoping call is free.
Free scoping call. You own the code.
Was this page helpful?

