You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[FEATURE]:Asynchronous user agent communication. #15969
For the scenario described (Agent needs to ask a question).there might be a solution that just uses 2 sessions
There's 3 distinct properties of each solution by complexity.
Recursiveness: Each session can theoretically spawn other session of the same type as the main session. An example would be the 'subagents' solution.
Arbitrary length sessions: A main session can spawn an unlimited amount of subsessions of a different type. Also some 'subagents' proposals.
Fixed length alternative sessions: At most one additional type of session is added or split. For example a solution where we have a session for the conversation, and another that
Obviously solutions of the first and second type increase the scope of the project, the complexity and cost of the project.
An acceptable solution would be of the third type, not adding recursiveness and not adding the capacity to add an arbitrary amount of additional context. If average token per context is increased, it should be an n+1 cost increase.
A hacky implementation of this (that anyone can try), is asking the model, either via prompt, skill, or AGENTS.md, to communicate through files, either by using a MESSAGES.txt or through comments.
The issue of this workaround is that the models are not quite trained to follow this instruction. And they usually just don't attempt to communicate or to check communications.
Activity
github-actions commentedon Mar 4, 2026
This issue might be a duplicate of existing issues. Please check:
TZubiri commentedon Mar 4, 2026
#5887
TZubiri commentedon Mar 4, 2026
For the scenario described (Agent needs to ask a question).there might be a solution that just uses 2 sessions
There's 3 distinct properties of each solution by complexity.
Obviously solutions of the first and second type increase the scope of the project, the complexity and cost of the project.
An acceptable solution would be of the third type, not adding recursiveness and not adding the capacity to add an arbitrary amount of additional context. If average token per context is increased, it should be an n+1 cost increase.
A hacky implementation of this (that anyone can try), is asking the model, either via prompt, skill, or AGENTS.md, to communicate through files, either by using a MESSAGES.txt or through comments.
The issue of this workaround is that the models are not quite trained to follow this instruction. And they usually just don't attempt to communicate or to check communications.
github-actions commentedon Jul 5, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.