Agent Plugins, stateless MCP, and ARD: the agent interoperability stack is separating into layers.

The Agent Plugins announcement frames the plugin as a directory with a small manifest and fixed locations for skills and MCP configuration. That makes a useful unit portable across compatible clients without forcing the skill or server itself into a vendor-specific wrapper. It also creates a clear object for software-supply-chain controls: a package can have a publisher, version, digest, review status, dependency record, and allowed client set.

Portability increases the importance of release discipline. A package that works in more clients can also spread a vulnerable server configuration, over-broad instruction, stale reference, or unsafe default more quickly. Enterprise adoption therefore needs signed or otherwise verifiable provenance where supported, pinned versions, staged promotion, revocation, and a record of which client executed which release.

ARD uses domain-hosted catalogs and federated registries so an agent can find a capability without every client maintaining a private hard-coded list. The specification also allows verifiable trust metadata to travel with a result before the client connects directly through the resource's native protocol. This is materially different from ranking tools by a natural-language description alone.

A registry result should enter a policy gate, not a tool call. The consumer needs to verify publisher identity, requested capability, allowed data class, jurisdiction, required assurance, current version, and whether a direct connection is permitted. Registry relevance can help shortlist a resource; it cannot approve expenditure, sensitive-data access, or a real-world action on behalf of the user.

The 2026-07-28 MCP update removes the protocol handshake and logical session identifier from the core request model. A request can be handled independently by available workers, while long-running work uses explicit task primitives and application-managed state. This is a better fit for autoscaling and failover, but it does not eliminate state. It moves identity, task progress, idempotency, retries, and recovery into services that teams must design and observe deliberately.

The update also sharpens security and lifecycle boundaries through issuer verification, resource indicators, richer JSON Schema support, and a formal deprecation process. Those controls are valuable because agent tool calls are delegated authority. A token intended for one MCP server should not silently become authority at another, and a client should not learn that a protocol feature disappeared only when production work fails.

For Neura Parse, the useful product connection is architectural. NowFlow is positioned as an agentic workflow and application layer with approvals, state, retries, tools, notifications, and reviewable work. That is the control surface around any capability package or protocol adapter: who requested the action, what policy applied, whether evaluation passed, who approved release, what the tool changed, and how the outcome was reviewed.

This article does not announce Agent Plugins, ARD, or MCP 2026-07-28 support in NowFlow. Those are separate product decisions that require implementation and verification. The near-term action is to design adapters and registries behind stable internal contracts so new specifications can be evaluated without rebuilding the authority model for every client or tool.

A credible pilot starts with one workflow whose user, tools, data boundary, approval policy, and rollback are already understood. Package one capability, publish or register it in a controlled environment, invoke it through a pinned protocol version, and replay the complete path from discovery to outcome. The goal is not the number of tools an agent can find. It is whether the organization can prove why a particular capability was selected and what authority it exercised.

Measure discovery precision, verification failures, authorization denials, evaluation outcomes, human interventions, task recovery, duplicated side effects, latency, cost, and rollback success. Those measures turn interoperability from a compatibility story into an operating decision. Only then should the catalog, client set, data access, or autonomy envelope expand.