[RFC] A coordination substrate for multi-agent OpenJarvis (messaging, dispatch, shared memory)

[gilbert-barajas]

(Discussion issue — per CONTRIBUTING, opening for discussion before any PR, since this proposes a new primitive touching BaseAgent / Tools & Memory. Deferential by design: this defers to your operator-composition direction and invites co-design.)

Motivation

OpenJarvis today is strongest at single, composable agents. The roadmap names the next frontier but marks it open:

• Operator composition / chaining — Design Needed ("operator A feeds results to operator B").
• Agent-to-agent coordination — no explicit messaging/dispatch layer today.
• Federated memory — synchronizing memory across devices — in focus.

As local-first personal AI grows past one agent, these converge into one missing piece: a coordination substrate — a place for agents to message each other, dispatch and track work, and share persistent memory + identity across sessions and devices. Building this well (durable queues, addressing, persistence, observability) is a real systems project on its own.

Proposal

Adopt a small coordination interface — in the spirit of how InferenceEngine abstracts serving across Ollama/MLX/vLLM — with Mycelium as the first pluggable implementation. Mycelium (SoftBacon-Software/mycelium, Apache-2.0, matching OpenJarvis) is an existing open coordination platform built for exactly this: tasks, memory, messaging, dispatch, savepoints, designed to run on hardware you own.

Mapping to the roadmap:

roadmap item	Mycelium primitive
operator composition / chaining	tasks + declared dependencies (A → B)
agent-to-agent coordination	channels / direct messaging + addressing
federated / persistent memory	memory backends + savepoints (cross-session identity)

Integration shape (respecting your constraints)

Per CONTRIBUTING's "may-not-be-accepted" list (no breaking changes, no heavy deps, no UX friction), this lands as an optional extra, not a core requirement:

• A CoordinationBackend interface registered via your existing registry pattern; NullCoordination (single-agent, today's behavior) is the default. Zero impact unless a user opts into multi-agent.
• The Mycelium backend is an optional install that talks to a Mycelium instance over its API (loose coupling) — no new heavy dependency in the core, and it keeps the license boundary clean (separate service).
• Start minimal: messaging + task dispatch first; memory/federation second, aligned with your federated-memory work.

What we bring

• Mycelium is live, open (Apache-2.0), and purpose-built for multi-agent coordination + persistence on local hardware.
• Gilbert Barajas (@gilbert-barajas) is an active OpenJarvis contributor — PRs chore: replace deprecated pynvml with nvidia-ml-py #375 / fix(engine): modernize apple_fm_shim for the public apple-fm-sdk #377 / fix(engine): convert apple_fm_shim cumulative snapshots to OpenAI deltas #379 / fix(server): preserve tool_calls when rebuilding streaming-chat history #383 / feat(server): expose sampler params (repetition_penalty / top_p / top_k / min_p / *_penalty) for managed agents #387 / fix(server): inject dependencies into managed-agent streaming tools #396 plus dogfood-found issues across the managed-agent-streaming, persona-loading, and apple_fm clusters — happy to own this as a sustained workstream (the path-to-maintainer domain).
• A reference multi-model "squad" (reasoner + coder + verifier, coordinated + persisted via Mycelium) we can adapt into an example preset.

Open questions (for co-design)

1. Should coordination be a core primitive or strictly an optional extra? (We lean optional-first.)
2. How should the interface relate to your planned operator-composition design — does composition sit above a coordination backend, or are they the same layer?
3. Where does this meet federated memory — one backend, or memory-as-a-Mycelium-capability?

Honest notes

• This is a proposal to fill a gap, shaped to your architecture — not a finished design. We'll build to your interfaces.
• Scope discipline: ship the smallest useful slice (messaging + dispatch behind a default-off backend) and earn the rest.