The A2A context problem: why multi-agent flows break at the handoff

Multi-agent (A2A) workflows are the next frontier for enterprise AI, one agent triages an incident, hands it to a remediation agent, which hands it to a deployment agent. The promise is powerful. The reality is fragile.

The failure point is almost always the same: context loss at the handoff boundary. Agent A completes its task and passes a task ID to Agent B. Agent B starts cold, no knowledge of what Agent A learned, what systems it queried, or what the current state of the infrastructure is. It rebuilds context from scratch, often from different (and inconsistent) sources. The compounding intelligence that makes multi-agent flows valuable never materialises.

This happens because most agent architectures are built around isolated context windows, not shared knowledge infrastructure. Each agent has its own retrieval pipeline, its own vector database, its own view of the world. Handoffs pass task parameters, not knowledge.

The fix requires two things working together: a shared knowledge graph that all agents query (so every agent starts with the same ground truth), and an A2A protocol that carries graph context with the task (so handoffs aren't just parameter passing, they're knowledge-aware transfers).

In Cendriix, the unified knowledge graph (Prism) is the shared foundation every agent queries. The A2A gateway uses signed agent cards, scoped short-lived tokens, and an explicit consent model for cross-agent handoffs. When Agent A hands off to Agent B, the graph context, the full resolved view of the infrastructure at that moment, travels with the task. Agent B starts where Agent A finished, not from zero.

For enterprise teams, this changes the economics of multi-agent workflows fundamentally. You're not paying for each agent to rediscover the same context. You're paying for compounding execution where knowledge accumulates across the full workflow chain.