On August 6, 2026, AWS Agent Registry leaves public preview. In the same move, AWS relocates it out of the bedrock-agentcore namespace — where it lived since its April preview — into a namespace of its own: agent-registry. Anyone already building against it has to update endpoints, IAM policies, SDK clients, CLI scripts, and registry data before the switch, per AWS's own migration guidance.
If you read one line: the layer where agents and tools get found just got a hyperscaler default, and a service that earns its own namespace is a service AWS intends to make load-bearing — so decide now where your agent becomes discoverable, because your buyers' tools are about to have an obvious first place to look.
What actually happened#
AWS Agent Registry is a managed catalog for agents, tools, MCP servers, agent skills, and custom resources. You publish a record, it goes through an approval workflow (draft, pending, approved), and then both humans and other agents can find it using hybrid search — describe a use case in natural language and semantic matching surfaces the right record even when the keywords don't line up. It shipped in preview in April to, in InfoQ's framing, "govern AI agent sprawl across enterprises."
The August 6 change is two things at once. The obvious one is graduation: preview to general availability. The quieter, more telling one is the namespace. Moving from bedrock-agentcore to agent-registry means AWS is no longer treating discovery as a feature bolted onto its agent runtime. It's a standalone primitive with its own IAM surface, its own endpoints, its own identity.
A service that gets its own namespace is a service the vendor plans to build a decade on top of.
The other piece that matters is the Registry MCP endpoint. The registry doesn't just serve a console for humans — it exposes itself as an MCP-compatible server (following the 2025-11-25 MCP spec) so an agent or an IDE can query it directly and discover other agents and MCP servers on its own. That's the important verb: agents discovering agents, programmatically, without a person in the loop.
Why the discovery layer matters#
For a year the discovery conversation lived one level down, at tools: where do I find MCP servers? The community answered it with a neutral, singular index — the open MCP registry. We walked through how these two catalogs differ, and why the stack is already fragmenting, in agent registry vs MCP registry for discovery.
The question has now moved up a layer, to agents, and the answers are arriving from cloud vendors instead of the community. That changes the physics. A neutral registry optimizes for everyone can be found. A hyperscaler registry optimizes for what my customers can safely use inside my platform — approval gates, audit trails, org boundaries. Both are legitimate. But they point discovery in different directions, and a founder has to know which one their buyer is actually querying.
Here's the non-obvious part: discovery is quietly becoming the new distribution. When an enterprise's agents search a registry to decide which tool to call, being in that registry — approved, well-described, semantically findable — is the modern equivalent of ranking in a marketplace. It's a channel, not a checkbox. We argued the general case in discovery is the new distribution; the AWS namespace move is that thesis turning into a shipping default.
And the timing rhymes with the protocol itself maturing. The final MCP spec shipped Apps and Tasks just days ago — the primitives that let agents do richer work. Registries are the primitives that let agents find that work. Capability and discovery are hardening in the same quarter, which is usually the sign a layer is about to get real defaults.
What a founder should do this quarter#
Concrete moves, in order of leverage:
- If you already use the preview, migrate before August 6. Update your endpoints, IAM policies, SDK clients, and CLI scripts to the
agent-registrynamespace. Treat it like any breaking API cutover, not a cosmetic rename. - Decide where your agent or tool needs to be discoverable. If you sell into AWS-heavy enterprises, being listed and approved in their Agent Registry is now part of the sale. If you sell to everyone, the neutral MCP registry still matters. Most founders need both.
- Write records for machines, not just humans. Semantic search rewards clear capability descriptions. The buyer running the query might be another agent typing "payment reconciliation," not a person browsing.
- Watch the other hyperscalers. AWS gave discovery its own namespace first. When a category leader draws that line, peers tend to follow — and a multi-registry world is a per-vendor listing tax you'll want to budget for.
The preview-to-GA line is easy to skim past as routine AWS housekeeping. The namespace is the signal underneath it. Agent discovery just stopped being a corner of Bedrock and became its own thing with its own front door — and the vendors are now competing to be the place your buyers' agents look first. Pick your registries before that choice gets made for you.



