Contents
Fast answer
AutoGen is free open-source software with no platform fee, seat licence or execution meter, and that is not the interesting part of its cost. As of 20 August 2026 the repository carries a notice that AutoGen is in maintenance mode, will not receive new features or enhancements and is community managed going forward, with only bug fixes, security patches and documentation improvements planned. The README directs new users to Microsoft Agent Framework, which it calls the enterprise-ready successor to AutoGen, and provides a migration guide. The repository showed 558 open issues when we checked. So the real cost of starting here in 2026 is two things: model tokens, which for conversational multi-agent runs are genuinely unpredictable because turn counts are emergent rather than specified, and a migration to Agent Framework that is a rewrite of the orchestration layer rather than a package rename. Neither has a published number, and only one of them is optional.
Facts last verified 2026-08
Quick facts
- Licence fee
- $0, free open source (checked 2026-08-20)
- Project status
- Maintenance mode, community managed
- What it still receives
- Bug fixes, security patches and documentation improvements only
- Microsoft's direction
- README tells new users to start with Microsoft Agent Framework
- Open issues
- 558 shown on the repository when checked
- Model tokens
- Provider rates; multi-agent runs multiply calls per outcome
- Token predictability
- Poor, because turn counts are emergent rather than specified
- Migration cost
- An orchestration rewrite, not a package rename
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.
- 01What it costs
- 02The real meter
- 03Worked example
- 04Maintenance mode
- 05Cheapest honest path
- 06When to leave
Why does this matter now?
A zero licence fee is the easiest thing in software to over-value. AutoGen's actual cost profile is a token bill you cannot forecast and a rewrite you will owe, and both of those are larger than any subscription in this directory. For an existing deployment that works, maintenance mode is survivable. For a new project, the vendor has told you in writing where to start instead, and starting anywhere else is a decision to pay for the migration twice.
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 why the licence fee is not the cost
- Budget unpredictable multi-agent token spend
- Price the migration you will owe later
- Decide against starting here in 2026
How do you do it, step by step?
1. Read the repository notice before the tutorial
The AutoGen README states plainly that the project is in maintenance mode, will receive no new features or enhancements and is community managed, and that new users should start with Microsoft Agent Framework, described there as the enterprise-ready successor. That single paragraph is more consequential to your total cost than any comparison of syntax.
2. Budget tokens as a range, not a number
Conversational multi-agent runs consume several times more tokens per outcome than a single agent, because the agents talk to each other. Turn counts are emergent rather than specified, so two runs of the same task can differ several-fold. Budget per completed outcome with a wide band, and cap turns explicitly in code rather than trusting convergence.
3. Cap the conversation before it costs you
The characteristic AutoGen failure is agents conversing at length without converging, which is expensive and produces nothing. Set hard maximum turn limits, add a termination condition tied to a concrete output rather than agreement, and log token consumption per run so you find the runaway cases in testing rather than on the invoice.
4. Price the migration you are signing up for
Microsoft publishes a migration guide from AutoGen to Agent Framework, and the successor moved away from AutoGen's conversational group-chat centrepiece toward typed workflows. That means the pattern you would learn here is the one Microsoft moved away from, and porting is engineering work on the orchestration layer rather than a dependency bump. Estimate that as real weeks.
5. Sandbox anything that executes code
Code execution by agents is a genuine security exposure and needs a proper sandbox rather than good intentions. This is a cost line too: the container, the isolation and the review that make agent-written code safe to run are part of what AutoGen does not give you for free.
A zero licence fee is not the same as a low cost
AutoGen costs nothing to license. No platform fee, no seats, no execution meter, and that number is genuinely zero rather than a free tier with a cliff behind it. It is also the least informative fact about what running AutoGen costs, because the two real expenses are a token bill you cannot forecast and a migration you have already been told you will owe. Neither appears on any pricing page, and both are larger than most subscriptions in this directory.
Start with the migration, because it is the one with a published cause. As of 20 August 2026 the AutoGen repository carries a notice that the project is in maintenance mode, will not receive new features or enhancements, and is community managed going forward, with only bug fixes, security patches and documentation improvements planned. The README states that new users should start with Microsoft Agent Framework, describing it as the enterprise-ready successor to AutoGen, and a migration guide is provided. The repository showed 558 open issues when we checked.
That is the vendor telling you in writing where to go. For an existing deployment it is survivable and not urgent: maintenance mode means security patches continue and working systems keep working. For a new project it changes the calculation entirely, because you would be learning a conversational group-chat pattern that Microsoft's successor deliberately moved away from in favour of typed workflows, and paying to unlearn it later at a moment you will not choose.
Price that migration as engineering weeks rather than an afternoon. It is a rewrite of the orchestration layer, not a package rename, because the abstractions changed rather than just the imports. Whatever number you put on it, it is a cost that a paid framework with a stable roadmap would not have imposed, which is the specific sense in which free software can be the expensive option.
The token bill you cannot forecast
The second real cost is model tokens, and AutoGen's design makes them unusually unpredictable. Conversational multi-agent runs consume several times more 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. The problem is that it is not, because conversation length is emergent rather than specified: the agents decide how many turns a task takes.
The practical 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 not a rare pathological case, it is the normal risk of the pattern, and it is why budgeting AutoGen as a single number per run is unsound.
Budget per completed outcome with a deliberately wide band, and then narrow it with engineering rather than with estimation. Set hard maximum turn limits in code. Add a termination condition tied to a concrete output artefact rather than to the agents agreeing with each other. Log token consumption per run from the first day so that runaway cases surface in testing rather than on an invoice. Those three changes convert an unbounded risk into a bounded one.
There is a security cost sitting alongside the token cost that is easy to leave off a budget. Code execution by agents is a real exposure and needs a proper sandbox rather than good intentions. The container, the isolation and the review process that make agent-written code safe to run are all part of what AutoGen does not provide for free, and they are engineering work with an ongoing maintenance tail rather than a one-off setup task.
What maintenance mode should actually change
The right response to maintenance mode is calibrated rather than dramatic, and the two situations are genuinely different. If you have a working AutoGen deployment, nothing broke on the day the notice appeared. Bug fixes and security patches continue. Migrating urgently is cost with no benefit, and the sensible plan is to schedule the move for the next time you are making substantial changes to that system anyway, so the rewrite rides along with work you were doing.
If you are starting something new, the calculation is not close. Microsoft Agent Framework is the sanctioned destination, is also free, and has a published migration guide from AutoGen, which tells you something about how seriously the vendor takes the transition. Choosing AutoGen for greenfield work in 2026 means adopting a pattern the vendor replaced and paying for the migration twice, once in learning and once in rewriting.
If the appeal was specifically the role-based multi-agent design rather than Microsoft, the alternatives are worth knowing. 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. That is the closest philosophical match. LangGraph takes the opposite approach, making control flow explicit as a graph, which removes the non-converging conversation failure class by construction rather than by turn caps.
And it is worth asking whether the problem needed multiple agents at all. Multi-agent designs multiply model calls per outcome, and a great many of them 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 easier to debug. If a single agent handles your task, that is the cheapest answer available and it does not require choosing a framework at all.
At a glance
Microsoft AutoGen cost profile, read from github.com/microsoft/autogen on 20 August 2026
| Line item | Published price or status | Note |
|---|---|---|
| AutoGen framework | $0 | Free open source, no platform fee, seats or execution meter |
| Project status | Maintenance mode | Repository states it will not receive new features or enhancements |
| Governance | Community managed | Going forward, per the repository notice |
| Ongoing changes | Bug fixes, security patches, documentation | No feature development planned |
| Recommended alternative | Microsoft Agent Framework | README calls it the enterprise-ready successor and provides a migration guide |
| Open issues | 558 | As shown on the repository when checked on 2026-08-20 |
| Model tokens | Provider rates | The entire recurring cost; not published by Microsoft because it is not Microsoft's meter |
| Hosting | Ordinary container cost | A small server or container to run the agents |
| Migration to Agent Framework | Not published | Engineering time; the successor uses typed workflows rather than group chat |
What mistakes should you avoid?
- Treating a zero licence fee as a low total cost
- Starting a new project on a framework whose own README points new users elsewhere
- Budgeting multi-agent token spend as a single number when turn counts are emergent
- Letting agents converse without a hard turn cap or a concrete termination condition
- Assuming migration to Agent Framework is a package rename rather than an orchestration rewrite
- Running agent-generated code without a sandbox
FAQ
- How much does Microsoft AutoGen cost?
- Nothing to license. It is free open-source software with no platform fee, seats or execution meter. Your costs are model tokens billed by whichever provider you use, a small server or container to run the agents, and the engineering time to build and operate the surrounding system.
- Is AutoGen still maintained?
- It is in maintenance mode. The repository states it will not receive new features or enhancements and is community managed going forward, with only bug fixes, security patches and documentation improvements. The README directs new users to Microsoft Agent Framework and provides a migration guide. The repository showed 558 open issues when we checked on 20 August 2026.
- Why is the token bill unpredictable?
- Because conversation length is emergent rather than specified. Agents decide how many turns a task takes, so two runs of the same task can differ several-fold in consumption. Multi-agent runs also multiply calls per outcome against a single agent. Cap turns in code and budget per completed outcome with a wide band.
- Should I start a new project on AutoGen in 2026?
- Microsoft's own answer is no. The README names Microsoft Agent Framework the successor and tells new users to start there. Choosing AutoGen for greenfield work means learning a pattern the vendor moved away from and paying for a migration later, which is a larger cost than anything you would save.
- What does migrating to Agent Framework involve?
- A rewrite of the orchestration layer rather than a dependency change. Microsoft publishes a migration guide, and the successor replaced AutoGen's conversational group-chat model with typed workflows, so the structure of your agent code changes rather than just its imports. Treat it as weeks of engineering, not an afternoon.
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?

