Why it matters
See the actual architecture pieces (planning, subagents, memory, storage backends) that separate agents which survive multi-step work from ones that fall apart mid-task.
The tokenmaxxing angle
Subagents that isolate work in their own context window are a direct token-spend lever — each isolated context avoids re-feeding the whole conversation history, which is where multi-step agent costs quietly balloon.
From the organizers
Justin Kaiser live-builds an agent from a single create_deep_agent() call, then adds web search, subagents, and persistent memory.