BreakingTech retrospective archive — August 24, 2026 event.
When a company experiments with two or three AI agents, it is easy to know where they are and what they are doing. When they become dozens or hundreds, a new problem emerges: discovery. On August 24, AWS introduced Agentic Resource Discovery, an open specification designed to make it easier to find agents, tools, skills, and MCP servers across different environments.
The proposal integrates with AWS Agent Registry, which serves as a governed catalog for an organization's agentic resources. The idea is similar to what package managers and service catalogs have done in other phases of software: preventing every team from rebuilding what already exists from scratch.
The problem is no longer building an agent
The tools to build agents are becoming increasingly simple. This lowers the barrier to entry but increases fragmentation. One department creates an agent to query the CRM, another builds one for internal documents, a third exposes an MCP server: without a central registry, no one has a complete picture.
The consequence is not just inefficiency. It is a security issue. A forgotten agent can retain access to sensitive systems, use obsolete credentials, or continue invoking tools that are no longer approved.
Discovery and governance become infrastructure
ARD attempts to define a common language for describing and finding agentic resources. The value of an open standard is enabling tools created in different environments to be discovered without relying on a single client or a single platform.
For enterprises, this means being able to build policies around agents: who can publish them, who can use them, what tools they can invoke, and how they are audited.
Agents increasingly resemble software employees
The more autonomy an agent is granted, the more an identity, an owner, and an authorization perimeter must exist. It is the same logic companies use to manage human accounts and cloud services, but applied to software capable of making decisions.
The next phase of enterprise AI could therefore be less spectacular than new benchmarks and far more infrastructural: registries, permissions, logs, versioning, and governance tools. This is what makes it possible to move from demo to large-scale deployment.
