Harmovela: Async Coordination Protocol Complementing MCP for Agent Systems

Introduction

MCP (Model Context Protocol) standardizes synchronous tool discovery and invocation for agents. However, when tasks span multiple agents, tools, memory systems, and runtimes over minutes or hours, systems need answers to questions such as: what stage is a task in? Who has accepted delegation? Is state still fresh? Which decisions become invalid after memory updates? Should execution retry, replay, or compensate after interruption? How does cancellation propagate downstream? How do multiple runtimes maintain consistent governance and audit semantics?

MCP vs. Harmovela

Harmovela is an open coordination protocol for autonomous systems. It complements MCP rather than replacing it. The division of labor is:

MCP handles synchronous capability discovery and invocation.

Harmovela handles asynchronous, incremental, replayable continuous coordination.

Seven Core Dimensions

Harmovela covers seven primary dimensions:

Event – describes what happened, including publish, subscribe, correlation, acknowledgment, and replay.

Task – describes what is executing, with lifecycle states: submit, accept, execute, progress, complete, fail, cancel.

State – describes what is true now, introducing versions, freshness windows, invalidation, and change propagation.

Context/Memory – describes which information influences decisions, including context updates, provenance, retrieval readiness, and memory invalidation.

Delegation – describes who is responsible for what, including task assignment, acceptance, transfer, escalation, and cancellation propagation.

Recovery – describes what to do after failure, including idempotency, retry, replay, checkpoint, interruption, and compensation.

Governance – describes who may do what, including identity, authorization, audit, tenant isolation, and policy integration.

Example Integration

An agent can invoke an external tool via MCP; Harmovela then carries the associated task progress, state changes, result streams, failure events, and recovery actions. Even when the actual tool is not an MCP tool, Harmovela can coordinate browsers, robots, IDEs, cloud runtimes, and multi-agent platforms.

Additional Modules and Infrastructure

The protocol also adds dimensional modules for Tool, Agent, Environment, Adaptation, Command, and Query, plus infrastructure modules: Harness, Runtime, CLI, Conformance, and MCP Bridge.

Implementations and Testing (Current Public Version)

Four production-grade implementations: TypeScript, Python, Go, Java.

Each includes Runtime Daemon, CLI, HTTP API, and MCP Bridge.

Seven transport bindings: stdio, WebSocket, SSE, gRPC, NATS, Kafka, Redis Streams.

Delivery stores: SQLite and PostgreSQL.

Retry, Dead Letter, and Replay mechanisms.

Cross-language consistency validation samples.

Approximately 700 tests.

Value Proposition

Harmovela's value is not in reinventing a message queue. Generic messaging systems transport data but do not understand whether a message represents task acceptance, state invalidation, delegation transfer, or compensation completion. The protocol layer must define semantics that autonomous participants share, not just how messages travel from A to B.

Emerging Agent Infrastructure Layers

Model APIs provide reasoning.

MCP provides synchronous capability invocation.

Harmovela provides cross-participant asynchronous coordination.

Concrete runtimes and business systems own execution and state authority.

Conclusion

As agents move from single tool calls to long-running, multi-party, cross-system execution, "coordination" becomes a distinct protocol problem, just like invocation.

Project repository:

https://github.com/axisrobo/harmovela