Snapshot
AI agent adoption is now shaped more by discovery and approval than by model comparison
Reading the official material available by April 27, 2026 across OpenAI, Anthropic, Microsoft, Google Cloud, and AWS shows that agent adoption is moving to a different layer. The main question is no longer only which model is smartest, which runtime lasts longest, or which protocol connects the most systems. Those still matter, but they no longer explain why an agent actually gets deployed. What matters more is where the agent is discovered, who is allowed to enable it, which actions can be approved, what metadata is available for comparison, and how the agent enters a managed work surface. This week is therefore best read not as a single launch, but as a convergence week in which app directories, stores, marketplaces, and registry layers became comparable public product surfaces.
25
Primary sources
Official docs and official announcements from five vendor ecosystems are enough to support the whole argument without relying on third-party coverage.
5 stacks
Surfaces compared
Consumer apps, workspace controls, enterprise registries, cloud marketplaces, and procurement agents are all now visible side by side.
4 layers
Adoption frame
Discovery, approval, metadata, and deployment now form a practical comparison frame for agent adoption.
1 shift
The question changed
The key question is shifting from “which agent is smartest?” to “which agent can be found, approved, and admitted into real work?”
Build-up
The late-April 2026 view sits on almost a year of accumulated work around distribution and governance surfaces
Anthropic moved Integrations into Claude’s main product story
Remote MCP became part of the default product narrative, linking app connections, research, and action in the same Claude surface.
Anthropic added a connector directory and made discovery itself a product surface
The story shifted from “you can connect tools” to “users can browse and choose which tools to connect.”
AWS carved out AI Agents & Tools as a dedicated Marketplace category
MCP and A2A support, deployment options, and pricing became comparison fields in the catalog itself.
Google Cloud linked Marketplace distribution to Gemini Enterprise deployment
Partner-made agents, enterprise governance, and A2A Agent Cards started to look like parts of one connected adoption path.
Microsoft published Agent Store as a central hub with explicit review surfaces
Capabilities, data and tools, security and compliance, certification, and publish flows all appeared as first-class elements of the agent experience.
OpenAI unified connectors into apps and made the directory plus action controls explicit
Search, deep research, sync, write actions, and custom MCP-backed apps now sit inside one user-facing Apps surface.
What makes this week notable is that these no longer look like isolated product details. OpenAI now shows the end-user app layer. Anthropic puts integrations and directories into ordinary Claude usage. Microsoft makes the approval chain and registry public. Google Cloud and AWS make partner distribution and procurement concrete. AI agents are becoming easier to compare as assets that must be admitted, not just as models that can execute a task.
Four Surfaces
Current agent adoption becomes easier to understand when it is split into discovery, approval, metadata, and deployment
1. Discovery
The first requirement is a place where agents can be found. OpenAI has the ChatGPT app directory. Anthropic has the Claude directory. Microsoft has Agent Store. Google Cloud and AWS have marketplaces. At this layer, search, categories, listings, and featured collections begin to matter as much as raw model performance.
2. Approval
Discovery is not enough; someone still needs the power to enable or reject the agent. OpenAI exposes app enablement, action controls, parameter constraints, and domain restrictions. Anthropic exposes owner-level organization connectors. Microsoft turns requests, publishing, and approval into an explicit workflow. Approval is now part of the product, not an afterthought.
3. Metadata
The comparison object has also changed. Teams now care about capabilities, write permissions, certifications, pricing models, deployment options, protocol support, and agent cards. Google Cloud making A2A Agent Cards part of the listing contract and AWS foregrounding MCP and deployment filters are the clearest examples.
4. Deployment
The last layer is how the agent actually enters work. OpenAI points toward ChatGPT conversations and custom apps. Anthropic points toward Claude and Claude Desktop. Microsoft points toward Microsoft 365 Copilot. Google points toward Gemini Enterprise. AWS points toward AgentCore Runtime and Gateway. Without this layer, the catalog remains ornamental.
Observation
At this stage, an agent is easier to understand as a deployable catalog asset with metadata and policy boundaries than as a prompt bundle or isolated endpoint.
Vendor View
The direction is similar across vendors, but each one emphasizes a different control surface
OpenAI unifies the consumer-facing app surface
Apps in ChatGPT brings search, deep research, sync, write actions, and custom apps into one vocabulary. The Apps SDK also ties approved apps to the ChatGPT app store and Codex distribution, while developer mode exposes a read and write MCP client surface. OpenAI’s distinctive move is that the distribution layer is visible inside the main end-user product.
Anthropic brings MCP-backed integrations into normal Claude usage
Claude can now connect to your world made integrations part of the main product story, and the Claude directory adds one-click discovery. The help center then makes pre-built connectors, custom connectors, and owner-level organization enablement operationally concrete. Anthropic’s strength is how clearly it translates protocol ideas into user-facing connection flows.
Microsoft treats the agent as a registry object with an approval chain
Agent Store brings Microsoft, partner, and organization-built agents into the same catalog, while Admin Center surfaces capabilities, data and tools, security and compliance, certification, and activity. Agent 365 adds agent IDs, registry, access control, visualization, and security on top. Microsoft’s framing is the strongest on managed admission and organizational visibility.
Google Cloud connects partner distribution to enterprise deployment through A2A
In Cloud Marketplace, the AI agent becomes a listing object and the A2A Agent Card becomes part of the listing contract. In Gemini Enterprise, Google-made, partner-made, and custom agents are presented as things you can discover and govern together. Google’s distinctive move is to link marketplace metadata directly to enterprise deployment.
AWS foregrounds procurement and deployment choice as part of the product
AWS Marketplace for AI agents and tools turns semantic search, category browsing, deployment options, pricing models, compliance filters, and MCP support into the buyer surface. Agent mode then uses specialized search, comparison, and evaluation agents to support procurement itself. AWS’s strongest message is that agent adoption starts before runtime, at the catalog and buying layer.
Separation
Being discoverable is not the same thing as being safe to deploy
A directory listing is not a trust guarantee
A listing makes an agent easier to try, but not automatically safe to use. Teams still need to inspect what data the agent can read, what actions it can perform, what domains it reaches, and what certifications or attestations are available.
MCP or A2A support is not governance
Protocol support improves portability and discoverability, but it does not answer who can approve, revoke, or constrain the agent. That is why action controls, owner enablement, publish flows, and marketplace review still matter.
Approval flow is now part of product design
The practical product difference is no longer only what the agent can do. It is also how quickly a team can review, publish, block, and reconfigure that agent. Admission speed and accident rate both depend on the quality of this layer.
Procurement and operation remain separate problems
A marketplace can reduce buying friction without solving day-two observability or write-path governance. After discovery, the next real question is whether the organization can safely operate the agent once it is connected.
Workflows
This comparison axis matters most when agents are treated as organizational deployment targets, not personal experiments
Workspace research and document work
- Admins can enable search-heavy apps first while restricting write-capable apps.
- Users then connect the approved apps from a directory and use them directly in conversation.
- The practical difference often shows up in app availability and policy controls, not base model quality.
Ready-made sales or support agents
- Catalog surfaces like Agent Store make it easier to start from a prebuilt workflow instead of building from scratch.
- The real design questions become deployment scope, data access, handoff behavior, and review flow.
- Weak approval paths can block rollout even when the underlying agent is technically capable.
Partner-made agent procurement
- Google Cloud and AWS let teams compare agents through pricing, deployment option, and protocol metadata before committing.
- Buyers can evaluate container versus API delivery, compliance posture, and where the agent can be registered.
- Procurement becomes tightly coupled to architecture instead of staying a separate business workflow.
Internal distribution of custom agents
- Apps SDK, custom connectors, Agents SDK, and marketplace-style metadata make internal agents easier to distribute consistently.
- The bottleneck shifts from coding speed toward publish rules, review, update discipline, and disable paths.
- A clear distribution surface also helps reduce shadow agents inside the organization.
Adoption
The right platform depends on where the friction sits in your adoption path
If discoverability is the bottleneck, the catalog surface matters first
When teams do not know what is available, app directories, stores, and partner marketplaces matter more than small model deltas. The clearest win comes from making options legible.
If approval latency is the bottleneck, admin controls become the real comparison axis
RBAC, action control, parameter constraints, publish workflows, and domain restrictions determine whether agents spread safely or stall in review. This layer often matters more than the quality of the underlying demo.
If partner ecosystem reach matters, marketplace metadata becomes strategic
Rich pricing, deployment, certification, and protocol metadata makes buy-versus-build decisions faster. This is where Google Cloud and AWS become especially legible.
If portability matters, evaluate admission policy together with protocol support
MCP and A2A improve portability, but they do not solve publishing, blocking, or approval by themselves. The real question is whether the portable agent also fits the organization’s admission model.
If agent competition is read only through base-model quality, it becomes easy to miss the real deployment bottlenecks. In the public material now available, the next comparison axis is not only which agent is smartest, but which one can be found, approved, and placed into work under real governance constraints.