Why agent orchestration is needed

A single AI agent handles a bounded task well: find a product, answer a question, place an order. But
complex commerce scenarios — “find a gift for a five-year-old under $40 with next-day delivery” —
require search, filtering, stock checks, an assessment of delivery terms and payment, all at once.

Agent orchestration solves that through division of labour: the orchestrator breaks the goal into
subtasks and delegates each of them to a specialised agent that knows its own domain well.

Orchestration patterns

Orchestrator (planner)
  ├── Search agent       → "find products in category X"
  ├── Filter agent       → "keep only next-day delivery options"
  ├── Recommender agent  → "pick the best on rating and price"
  └── Checkout agent     → "place the order via ACP/AP2"

There are two main approaches:
– Centralized: one orchestrator drives every agent — simpler, but less resilient to failure
– Decentralized: agents interact directly over protocols such as A2A — harder to build, but it scales

The role of standards: A2A and MCP

Standard What it does
MCP Agent ↔ tools: connection to search, catalogue APIs, databases
A2A Agent ↔ agent: exchanging tasks and results between agents

Without standards, every integration is manual work. With them, any orchestrator can work with any
executor agent over a single interface.

Commercial applications

In agentic commerce, orchestration is what makes complex personalised purchase journeys possible. The
shopper’s own agent — inside ChatGPT or Perplexity, for example — acts as the orchestrator and
delegates tasks to the agents of specific stores or services, provided those support the standard
protocols.

Important: reliable orchestration needs error handling — one broken sub-agent must not take the
whole workflow down with it. The human-in-the-loop pattern adds a checkpoint before critical
actions such as irreversible payments.