AWS Agent Registry: runtime semantic discovery of MCP tools and A2A agents
AWS ML Blog · rss · 2026-10-21
The problem: hardcoded tools don't scale
Hardcoded tools work in demos, but enterprise systems expose hundreds of internal APIs (pricing, support, orders, inventory), so models burn tokens disambiguating tool names; every new capability forces orchestrator code changes, redeploys, and coordinated team releases.
Solution: Discover → Instantiate → Execute
AWS Agent Registry is a managed catalog for MCP servers, agents, and agent skills. The core loop:
- Discover: the orchestrator queries the Registry data plane in natural language, returning record identity, descriptions, and protocol descriptors (MCP server metadata or A2A agent cards) — but no invocation credentials, which must come via a separate trusted configuration path.
- Instantiate: build an MCPClient (pointing at a Bedrock AgentCore Gateway endpoint, authorized via Cognito OAuth) or an A2A client (invoking AgentCore runtime via IAM-signed boto3 calls) from the descriptor.
- Execute: an ephemeral inner agent is created with only the dynamically discovered tools; the LLM picks the right ones, and everything is torn down after the turn. Each request gets a fresh search and a fresh tool set.
Reference architecture
The sample deploys an order-management MCP server (AgentCore Gateway + Lambda), two A2A agents on AgentCore runtime, an orchestrator, and the Registry, using Bedrock (Claude Sonnet), Cognito, and Secrets Manager. Three demos: MCP-only (order status), MCP + A2A chaining (pricing/discounts), and pure semantic matching ("return and refund" matched an agent whose name contains neither word). The repo ships deployment, orchestrator, helper, and cleanup scripts.
Tip: write descriptions in user-facing language that distinguishes each capability from neighbors, and evaluate search quality with representative queries — catalog growth alone doesn't improve ranking.
More from coding & agent
- UPenn paper routes weak models to strong ones via model disagreement, no labels needed — pappasg69 · 2026-10-21
- Pat Hanrahan's 'On Notation' talk sparks a curated DSL course resource guide — blelbach · 2026-10-21
- General Video Agent enters public beta, edited its own launch video — ycombinator · 2026-10-21
- Beat AI's samey frontends: import a design system before generating UI — dotey · 2026-10-21
- FreePi Bets on Ad-Supported Coding Help to Undercut $20/Month AI Coding Subscriptions — heyneighbor · 2026-10-21
- RedMonk: MCP and A2A Verify Publishers, Not Safety — Registries Become the Trust Layer — rseroter · 2026-10-21