Skills are real for the swarm agent
Full skill playbooks now reach the model mid-run, loaded from a reviewed allowlist (PR
After this entry you'll know what a skill is and why the swarm agent just got noticeably better at repeatable tasks. A skill is a written playbook (a SKILL.md file) the swarm agent can pull in mid-run when the task matches, for example raise-a-pr or security-self-review. Until now skills were a facade: only the name and description reached the model, and the body, the actual instructions, was dropped.
What changes for you
The swarm agent now loads full skill bodies on demand, on planning, execute and resume paths, so runs follow the house playbooks instead of improvising. Skills come only from an explicit, PR-reviewed allowlist baked into the image:
skills: ["swarm-skills:raise-a-pr", "swarm-skills:use-hive-context", ...]A cloned repo cannot smuggle its own skills into a bee: an unlisted skill is rejected at invocation, which was verified live. If you want the swarm agent to do something a consistent way, the answer is now "write a skill"; see skills and MCP.