P11: Subagents and Context Isolation
What You Do
Run the same work two ways — one single conversation vs. isolated child conversations plus a synthesis step — and measure quality, tokens, cost, wall time, and compaction. Decide, with numbers, whether a branch earned its own context window. Three task shapes stress the boundary: a small repo audit, a breadth-first research corpus, and a large-repo investigation.
Harness Mechanism
Manual RemoteConversation children give per-child cost, tokens, wall time, and compaction so the context boundary is measured directly. A companion runner uses native OpenHands delegation (DelegateTool) traced in Laminar to show the real product surface. Child model routing, parallelism, and probe delay are knobs for testing why isolation might help.
Open First
Keep
A "when to use subagents" decision rule, grounded in your own table: what kind of task earns a separate context, what kind does not, and which cheaper alternative you would try first.
The reusable artifact is not the child-agent code. It is the habit of treating the context boundary as a harness decision that needs evidence — one you settle with your own numbers rather than a prior belief about whether subagents help.