Running several AI agents at once is easy. Trusting what comes back is not. This course takes parallel subagent work apart one seam at a time and asks, at each seam, what refuses a bad result. You will measure the serial fraction that caps every speed-up, give each agent a disjoint set of files to write, and keep one model resident behind a server that several agents share instead of loading it cold for every call.
Then you build the parts that say no: a reducer that gives the same answer in every arrival order, a receipt without which no verdict is Green, a quorum in which one FAIL blocks the merge, and a ratchet whose baseline never regresses. Every demo is a Rust program with pinned inputs, machine-checked assertions and a receipt. None of this makes an agent correct. It makes a wrong result stop the line before it ships.
