Agent Orchestration: When Multiple Agents Tackle One Task
A user asks their AI assistant: "Find me the best deal on a new sofa, buy the one with the best reviews, and schedule delivery for next Saturday." That single request might involve three or four different agents working together, each with different capabilities and different expectations from your website.
How Orchestration Works
An orchestrator agent receives the top-level task and breaks it into subtasks. Each subtask gets assigned to a specialised agent. A typical decomposition might look like this:
- Research agent: Visits furniture sites, extracts product details, compares prices and reviews.
- Decision agent: Analyses the collected data, ranks options, selects the best match.
- Purchasing agent: Handles the checkout flow, enters shipping details, processes payment.
- Scheduling agent: Coordinates delivery timing, checks calendar availability.
These agents may run sequentially or in parallel. The research agent might visit five different stores simultaneously, each in its own browser instance. The purchasing agent only activates once the decision agent has chosen a product.
Different Agents, Different Needs
Each agent type interacts with your site differently.
Research agents are focused on information extraction. They want to find product names, prices, specifications, reviews, and availability as quickly as possible. They favour structured data, clean HTML, and pages that load fast. They may visit dozens of pages on your site in quick succession. If your product pages have clear Schema.org markup, the research agent can extract what it needs without parsing your visual layout at all.
Purchasing agents need to complete transactions. They fill in forms, select options, handle authentication, and process payments. They need predictable UI flows, clear error messages, and forms that work without JavaScript quirks. A purchasing agent that encounters a complex checkout flow may abandon the transaction entirely.
Comparison agents visit multiple sites in parallel and need consistent data formats. If one site lists prices including VAT and another excludes it, the comparison breaks. Structured data with explicit currency and tax information helps these agents make accurate comparisons.
The Handoff Problem
When one agent finishes its work and hands off to the next, information needs to transfer cleanly. The research agent produces a summary: "This sofa from StoreName costs £899, has 4.7 stars from 312 reviews, and is available for delivery." The purchasing agent receives this summary and needs to find the exact product on the site.
This is where things often go wrong. The purchasing agent might land on the homepage and need to search for the specific product. If the product URL was passed along in the handoff, it can navigate directly. If not, it needs to search, which introduces another set of interactions with your site.
Stable, shareable product URLs matter. If your product pages use clean URLs that do not expire or change between sessions, the handoff between agents is straightforward. Session-dependent URLs, URLs with authentication tokens, or pages that require a specific referrer to load properly create friction.
Parallel Visits from the Same User
Orchestrated agents may send multiple simultaneous requests to your site. A research agent comparing five products might open five browser tabs at once. This looks like rapid automated traffic from the same IP address, which some rate-limiting systems will flag or block.
The distinction between malicious scraping and legitimate agent-driven shopping is difficult to make from traffic patterns alone. Both look like rapid, automated page requests. The difference is intent: one is extracting data to resell, the other is acting on behalf of a customer who might spend money on your site.
How you handle this traffic has real commercial implications. Block the research agent, and you never make it into the comparison results. Let it through, and you might gain a customer.
What the Orchestration Layer Tracks
The orchestrator keeps state across all its sub-agents. It knows which sites have been visited, what prices were found, which agents succeeded, and which failed. This accumulated knowledge influences future decisions. If your site consistently fails at the checkout stage, the orchestrator may learn to deprioritise you in future comparison tasks.
This creates a feedback loop. Sites that work well with agents get recommended more often. Sites that cause agents to fail get dropped from results. The effect is similar to how search engine rankings reward accessible, fast, well-structured sites.
Preparing for Multi-Agent Visits
The most practical thing you can do is make your site work well for a single agent. If one agent can browse your products, extract prices, and complete a purchase, orchestrated multi-agent systems will work too. The fundamentals do not change: semantic HTML, clear labels, fast load times, stable URLs, and structured data.
Agent-to-agent communication protocols are still emerging, and the specific orchestration frameworks will continue to evolve. But the underlying requirements from your website remain constant. Make your content easy to extract, your forms easy to fill, and your checkout easy to complete.