A parent agent investigating a regression delegates dependency research to one child and test-history analysis to another. Both finish, and the parent uses their results to choose the patch.
The children are not merely two tasks running in parallel.
The parent splits attention, not responsibility
Each child receives a bounded objective and context, executes through its own Session or Run lineage, and returns a result to the parent. The relationship records who delegated, which child performed the work, and where the result re-enters the parent's loop.
Depth and concurrency ceilings prevent an unbounded tree of children. Budget, tool, sandbox, and external-effect authority must be inherited or independently granted according to policy; a parent cannot delegate power it does not have.
A child needs enough source material and constraints to complete its objective, not the parent's entire conversational history by default. Bounded context reduces cost and accidental authority transfer. Result fan-in should preserve provenance so the parent can distinguish two child findings and resolve disagreement.
Three kinds of parallel work
A child result is evidence, not automatic acceptance. The parent decides how it changes the next step, whether another tool is needed, and what can be claimed in the terminal result.
A delegated child has agent lineage and returns evidence for parent synthesis. A task graph can describe the durable plan that motivated delegation, but it does not itself spawn a child. A parallel tool batch executes independent calls inside one loop step without creating a new agent relationship.
The agent loop shows how child evidence changes the parent's next decision. Task graphs and work state owns the durable plan around it.