Archon wraps Claude Agent SDK workflows

Archon’s real bet is that the winning layer in AI coding is the workflow harness, not another standalone agent. That’s a sensible direction: it makes the model interchangeable while keeping the process deterministic.

• –YAML workflows make the code-review, testing, and PR loop explicit, which is where most agent demos fall apart in real teams
• –Git worktree isolation is the right primitive for parallel agent runs because it limits collisions and makes failures easier to contain
• –Claude Agent SDK hooks and scoped skills give Archon a path to reusable playbooks, not just one-off prompt chains
• –The rewrite away from the older Python task-management/RAG version signals a clearer product thesis: orchestration over memory
• –The tradeoff is dependency risk, since harnesses like this inherit Anthropic SDK and billing changes before most users feel them