The harness determines what your orchestrator agent never sees, not the model.
I think subagents become more than context engineering once they can act.
Context should bound what a child knows; authority should independently bound what it can change. Otherwise we've isolated the reasoning without isolating the blast radius.
100% agree, but that's still not context engineering? 🤔
I think subagents become more than context engineering once they can act.
Context should bound what a child knows; authority should independently bound what it can change. Otherwise we've isolated the reasoning without isolating the blast radius.
100% agree, but that's still not context engineering? 🤔