Each delegated branch starts focused
Public Preview

Keep context across long-running work

Duale AI keeps long-running work coherent by separating stored task history from the active context sent to a model.

Duale AI keeps long-running work coherent by separating stored task history from active model context, supporting targeted recall and delegation or conversation continuation.

  • Child executors maintain separate working contexts and return outcomes as tool results to the parent.
  • Stored history, active context, and targeted recall serve distinct roles during long-running tasks.
  • Targeted recall retrieves selected original detail after older active context has been reduced.
  • Delegation handles focused work during a task; response.next() adds a turn after success.
  • Deadlines bound task duration but do not preserve context.

Summaries were generated by AI. Generative AI is experimental.

Duale AI keeps long-running work coherent by separating stored task history from the active context sent to a model.

Each delegated branch starts focused

A child receives a focused instruction and maintains its own working context. The parent keeps its context and records the child’s returned outcome as a tool result. Their full transcripts remain separate.

This separation keeps each executor focused while the parent receives the result it needs.

Stored history and active context are different

Duale AI uses three related views during long work.

Context viewPurpose
Stored historyKeeps original task messages and tool results available within the task history
Active contextGives the current model call the bounded information it needs
Targeted recallLets the model retrieve selected original detail after older active context has been reduced
  • Context view
    Stored history
    Purpose
    Keeps original task messages and tool results available within the task history
  • Context view
    Active context
    Purpose
    Gives the current model call the bounded information it needs
  • Context view
    Targeted recall
    Purpose
    Lets the model retrieve selected original detail after older active context has been reduced

The platform stores task messages and returned child outcomes before it continues the task. Later model calls can use a reduced active context and recall selected detail from stored history.

Long-running work explains reduction and targeted recall.

Stored history provides continuity. The deadline bounds the task’s duration; it does not preserve context.

Choose delegation or conversation continuation

Delegation handles focused work during a running task. response.next() adds a new turn after a successful task.

Decision pointDelegation inside a taskSDK conversation continuation
When it happensWhile a task is runningAfter a task finishes successfully
What it addsA child executor for focused workA follow-up task on the existing root executor
Who receives the outcomeThe executor that delegated the workYour application
Time boundaryNo later than the caller’s remaining deadlineNo later than the inherited conversation deadline
  • Decision point
    When it happens
    Delegation inside a task
    While a task is running
    SDK conversation continuation
    After a task finishes successfully
  • Decision point
    What it adds
    Delegation inside a task
    A child executor for focused work
    SDK conversation continuation
    A follow-up task on the existing root executor
  • Decision point
    Who receives the outcome
    Delegation inside a task
    The executor that delegated the work
    SDK conversation continuation
    Your application
  • Decision point
    Time boundary
    Delegation inside a task
    No later than the caller’s remaining deadline
    SDK conversation continuation
    No later than the inherited conversation deadline

Multi-turn continuation defines the SDK contract for response.next(). Executors and deadlines explains the work tree that delegation creates.

Use delegation when a running task needs focused child work. Use conversation continuation when your application needs a new turn after a successful outcome.