Use Git to share instructions and state across personal AI assistants
Git is probably the best way to share instructions and state across multiple personal AI assistants, since it provides an interface that pretty much all agents can use. It's also flexible (text files), and keeps history.
Personal AI assistants update and make changes to shared state using a set of instructions. Obviously, agents should not be able to change their own instructions, because bad instructions change all agent behavior. Instructions include things like AGENTS.md, skills, rules, playbooks, etc. State includes things like task logs / ledgers, memory, handoffs, agent communication, etc. Memory and notes should become agent instructions after human review. Agents should also be able to propose changes to their instructions.
Make things simple by using the repo-root instructions and skills instead of global ones. Sandboxed and runner agents generally work in a single repo, and it's harder to add global instructions there. Try to keep everything in a single repo, with GitHub push rulesets to prevent agents from modifying instructions.