AWS shows how AI assistants can find their tools as needed

AWS has shown a way for an AI assistant to find its tools at the moment a request arrives, instead of carrying every tool built in from the start. The approach uses AWS Agent Registry, a managed catalog. AWS says new tools become usable once they are registered and approved, with no code change or redeploy.

The problem it targets

AWS says hardcoded tools work for a handful, start to degrade at a few dozen, and become unmanageable beyond that. Large companies can have hundreds of internal systems for pricing, support, orders, inventory and customer data. The AI model can end up spending more effort telling tool names apart than working on the actual request. Each new tool also needs a code change, a redeploy and a release coordinated across teams.

How the lookup works

The registry catalogs MCP servers, agents and agent skills. MCP servers are services that offer tools to an AI assistant. The assistant sends a plain-language question to the registry. It gets back matching records with a name, description, type, status, version and connection details.

The assistant then builds a connection for each match and creates a temporary helper that holds only those tools. The AI model picks the right ones, and the answer goes back to the user. After the turn, the helper is removed. The next request starts with a fresh search.

Credentials stay separate

Search results do not include passwords or login methods for the tools. The application must get those through a separate, trusted path. In the sample, one kind of tool uses an access token fetched on its own. Calls to the other kind are signed with the caller's AWS credentials. The registry also only searches records that are approved and discoverable.

What the sample shows

The sample sets up an order-management tool server, two agent-to-agent (A2A) agents, a registry and an orchestrator that coordinates them. It runs three demos:

• Order status, which needs only the tool server.
• Pricing and discounts, which chains the tool server with an A2A agent.
• Return and refund, where those words are not in the matching agent's name.

The third demo tests whether search can match meaning rather than exact names. The material gives no speed, cost or accuracy numbers, so real-world savings are not yet shown. It is a reference sample, not a report from a live business.