Agents running in parallel on shared tasks have no native mechanism to share context, route information, or coordinate dependencies without human operators acting as manual routers. This creates a human-in-the-middle bottleneck that scales poorly as agent count increases, negating the throughput gains of parallelization. Existing frameworks treat agents as isolated execution units with no shared state primitive.

Parallel agents working on shared tasks can't exchange context, route dependencies, or coordinate state without a human manually copy-pasting between them — creating a bottleneck that defeats the purpose of parallelization.

AI engineers and agent framework developers (LangChain, CrewAI, AutoGen users) orchestrating 3+ agents on complex workflows like research, code generation, or data pipelines.

Teams already pay for orchestration tools (Temporal, Inngest) and agent frameworks (LangChain Plus, CrewAI Enterprise); a lightweight coordination primitive that plugs into existing stacks solves the exact gap these tools leave — shared state across agent boundaries — and the pain intensifies with every additional agent deployed.

MVP is a hosted pub/sub service with a Python/TS SDK offering three primitives: shared context channels, dependency signals (agent A blocks until agent B publishes artifact X), and a lightweight conflict resolution protocol; build on Redis Streams or NATS for sub-millisecond routing, with an OpenAPI-compatible webhook bridge so any agent framework can integrate in <10 lines of code.

The AI orchestration and agent infrastructure market is estimated at $2-5B by 2026; AgentBus targets the coordination layer within this, comparable to how Kafka carved a multi-billion dollar segment from broader data infrastructure.