Investigate support for long running Agent executions

[markwallace-microsoft]

Agents may require time to complete an execution e.g. A2A tasks or use of reasoning models. Review the design to ensure long running tasks can be supported.

• MEAI: Add a model for long-running operations
  • Add a model for long-running operations and apply it to the OpenAI Responses chat client.
    add support for background responses dotnet/extensions#6854
  • Add UserAgent header to the GetResponse{Streaming}Async methods
    • OpenAI SDK: Add OpenAIResponseClient.GetResponse{Streaming}Async overloads that take RequestOptions openai/openai-dotnet#753
    • Update to use internal GetResponse{Streaming}Async overloads
  • Update to use ResponsesContinuationToken instead of object. .NET: Change type of ContinuationToken properties dotnet/extensions#7050
  • Add the TextInputRequestContent and TextInputResponseContent classes to represent text input request and response used by A2A agents - Types to represent user input request and response for additional details dotnet/extensions#7088
• AF: Add a model for long-running operations
  • Add a model for long-running operations and apply it to the ChatClientAgent - .NET: Add support for background responses #1501
  • Support .NET 10
  • Update to the latest MEAI packages
  • Update to use ResponsesContinuationToken instead of object - .NET: Change type of props representing continuation token #2430
• AF: Add support for long-running operations to the A2A agent- .NET: Add support for background responses to A2A agent #2381
• AF: Add support for input requests to the A2A agent.
• AF: Add polling to the A2A agent.

[ekzhu]

One idea is to implement an agent using a workflow API which supports durable execution. When I trigger an agent that runs for a long time, the first thing I want to make sure is that failures and crashes can be recovered gracefully from the most recent self-consistent state.

[markwallace-microsoft]

Some additional context from @stephentoub

Some agents can have very long-running tasks. We will need the ability in either an IRunnableAgent or a separate IResumableAgent to resume consumption of a previously invoked operation. This will need to compose with orchestrations, such that an agent representing an orchestration over such resumable agents will itself be resumable.

[markwallace-microsoft]

Some additional context from @westey-m

We need to investigate how we can support long running processes.
This should include:

Supporting a way for the agent to return an id of a long running process.
Rejoining the long running process to get progress messages after it was started.
Waiting for the result of the long running process.
We need to decide whether a separate api is required to the regular Run apis. E.g. RunLongRunning, where the caller decides what API experience they want. Alternatively the current API can be updated to accommodate returning an ID if the agent decides it's appropriate.

[SergeyMenshykh]

ADR: #422
POC: #464