An agent’s model may select actions, but the application owns state persistence, authorization and the execution record. Single- and multi-agent architectures share these operational requirements. Parallel specialists introduce another problem: deciding which context and artifacts each participant should receive.
In this chapter
- Tool-using agents, state and guardrailsValidate arguments, persist steps and prevent unsafe duplicate effects.
- MCP, A2A and agent orchestrationSeparate tool integration, delegation and durable workflow execution.
- Shared agent context, handoffs and durable artifactsShare versioned evidence and structured state without duplicating entire conversations.
- Prompt injection, tool authorization and agent recoveryUntrusted evidence, permissions and idempotent recovery.