adrian7411374113/agent-team-brain

Agent Team Brain is a JavaScript based operating system for organizing teams of AI agents around a shared, file backed structure instead of one long chat transcript. Rather than treating each agent as a one off task runner, it gives every role, such as Coordinator, Architect, UX Reviewer, Scout, Analyst, Builder, and QA Reviewer, a persistent profile stored as markdown, along with its own workspace for scratch notes and current focus. Work moves through a defined loop: a task is created, sized for risk, assigned to the right role, tracked for blockers, and documented through shared artifacts like project briefs and handoff notes. Once built, work passes through a QA gate before being closed out. After a task finishes, the system captures an after action review and a trace score covering things like clarity, execution, and coordination quality, then rolls these into lessons that can update a role's playbook over time. A notable feature is the dream cycle, a way to compress and preserve context so long projects do not depend on an ever growing chat history. Individual agents produce short summaries of their work, which a coordinator consolidates into project state files, decision logs, and a context pack. This lets a fresh session pick up where a previous one left off by loading the role profile, playbook, and context pack instead of full history. The system also supports rotating out overloaded runtime sessions while keeping the durable team identity intact in files. You install it by running a CLI bootstrap command from Node, which scaffolds config files, role definitions, workspaces, and templates into a target directory. A doctor command checks the setup and reports pass, warn, or fail results. This is aimed at people building or coordinating multi-agent AI systems who want a repeatable, auditable structure rather than ad hoc prompting.