github/spec-kit reached 1.0.0 after a year, and the idea it argues for is simple: write down what you want built before any agent starts typing. It ships a process with commands for drafting, clarifying and planning, and it stays neutral about which assistant executes the spec, so Claude Code, Copilot and Cursor all work. Worth trying on the next feature that you would otherwise describe in a paragraph of chat.
If you run more than one coding agent at a time, stablyai/orca gives each one its own git worktree and puts them in a single view. You can fan one prompt across several agents, compare what each produced, and merge the one you like. It also has a phone companion that notifies you when a run finishes, which matters more than it sounds when a single task takes twenty minutes.
On the teaching side, THU-MAIC/OpenMAIC builds a course from a prompt and now adds a workbench where you steer the result. You can upload your own documents, audio and video and have it generate slides, quizzes and interactive pages from them, and runs survive a restart so you can come back to a half-built curriculum. It brings your own models and storage rather than locking you to one provider.