// the take
Hivemind is runtime infrastructure for multi-agent systems that sits underneath orchestration frameworks like LangGraph, giving them durable run/step/message/tool-call persistence, SSE event streaming, and an admin console. It's for teams who've outgrown a framework's local composition model and need an inspectable, operable service around their agents rather than ad-hoc logging and polling.
The adapter contract is genuinely minimal — one async method taking an AdapterContext and returning an AdapterResult — so swapping or mixing orchestration frameworks doesn't touch the API, schema, or event stream. Postgres as the single source of truth with Redis treated as purely ephemeral coordination (streams + pub/sub, no durable state) is the right call for a system that needs to survive worker restarts. The operational surface is unusually complete for a 1k-star project: tenant-scoped RBAC over both API keys and OIDC, HMAC-signed webhooks with retries, retention/purge endpoints, and audit events are all present, not just promised in docs.
The split between a Java/Spring Boot API and Python workers talking over a hand-rolled Redis job protocol means every change to run/event shapes has to be kept in sync by hand across DTOs, JPA entities, and Pydantic schemas — there's no shared schema/codegen layer visible, so drift between the two sides is an easy way to get silent production bugs. The README advertises AutoGen and CrewAI adapters as pluggable, but only LangGraph, Echo, and MCP adapters actually ship in backend/app/adapters — multi-framework support is a registration point, not working code you can use today. There's also a Temporal client/worker/workflow module living alongside the custom Redis queue and cancel-registry, which reads as two different job-execution mechanisms coexisting rather than one settled design — worth understanding which one you're actually supposed to build on before depending on it. Redis Streams gives at-least-once delivery, but the adapter contract shown has no idempotency key or dedup mechanism, so adapters that aren't naturally idempotent (anything issuing real side-effecting tool calls) can double-execute on redelivery.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →