AgentZero

AgentZero is a meta-agent - an agent whose job is to design other agents. It bootstraps new multi-agent project workspaces (top-level agents, sub-agents, skills) and then reviews, simulates, and scores them. It is intentionally architecture-aware: it makes the three-level decomposition decisions itself rather than punting them to the user, then closes the loop with a post-scaffold review and a simulation pass that runs the generated project and grades the transcripts.

Python 3.12, Poetry-managed, and multi-harness from day one - the same project artifacts run under either Claude Code or OpenCode via an adapter layer, not locked to one vendor. The repo carries the workflow primitives behind the scaffolding (/new-project, /add-agent), a two-tier skill-authoring template model, tenant worktrees for multi-org work, and a structured simulation-scoring rubric. Built as a tight Feb–Mar 2026 exploration of what self-scaffolding agent infrastructure looks like in practice.