A-Society structures how AI agents work on a project – a shared memory layer, explicit roles, and an enforced workflow – instead of one agent improvising every task with no memory between sessions.
What it does
-
Structured memory. It creates an
a-docs/folder inside your project: roles, rules, indexes, and project facts that agents read every session, so they start oriented instead of cold. - Roles + workflow. Work moves between roles through machine-readable handoffs, with records and closure checks – completeness comes from the process, not one agent deciding its own work looks done.
-
Self-improvement. After a flow, each role reflects on what happened and writes findings back into the project’s
a-docs/. - Cross-project feedback. Optionally those findings can be distilled into a report that improves the reusable templates new projects start from.
Every project gets an Owner role, and every flow starts and ends with it. The Owner takes in what you want to do, routes the work to the right roles, and is the only role that can close a flow – once every touched surface is accounted for.
Try it
curl -fsSL https://a-society.dev/install.sh | bash
Enter fullscreen mode Exit fullscreen mode
Node ≥ 18. Then configure a model in the UI — Anthropic or any OpenAI-compatible API — and start a flow.
Where it’s at
Still early, and the tooling around it is thin. Mainly I want feedback on whether the structure actually helps or is just overhead.
Repo: https://github.com/KartikGS/a-society
Site + docs: https://a-society.dev
답글 남기기