Swarm & multi-agent systems

make several agents work as one system — and know when not to

The Agent frameworks track answers "which framework, and how do I author an agent in it". This one answers the question that comes after: what happens when there is more than one. Multi-agent is sold as a way to get more done and is usually a way to get the same thing done less predictably. So the track starts with the primitive that actually matters — handoff, the moment one agent transfers control and state to another — and with the honest test for whether you need a second agent at all. From there: the interoperability layer, where you will learn to sort a new "agent protocol" into adopted, announced, or dead before you build on it; and fleets, where several coding agents work one repository at once and the bottleneck quietly moves from writing code to reviewing it. It ends where the field currently ends. Multi-agent evaluation tooling barely exists in 2026, so the last thing you learn is how to score a trajectory rather than an answer, and how to report variance across runs instead of the one run that worked.