June 2026 briefing: agents get enterprise identity, telemetry gets columnar, and the FDA codifies change control

The month in one paragraph

The Model Context Protocol declared its Enterprise-Managed Authorization extension stable, moving agent tool access under the corporate identity provider¹. At the end of the month, beta SDKs arrived for the stateless 2026-07-28 spec². OpenTelemetry published OTel-Arrow Phase 2 benchmarks, deprecated OpenCensus compatibility and made a pointed case against wrapping its API³⁴⁵. Kubernetes spotlighted the group behind Dynamic Resource Allocation and wrote up its policy on AI-assisted contributions⁶⁷. And in healthcare, the FDA codified a Class II pathway for machine-learning imaging software with a predetermined change control plan⁸.

MCP gets an enterprise identity story

Anyone who has rolled out agent tooling inside a company knows the friction: every user authorising every server one by one, no central policy and no single audit trail. The MCP project said as much, describing per-user authorisation as one of the biggest pain points in enterprise environments¹.

Enterprise-Managed Authorization makes the organisation's identity provider the decision-maker. During single sign-on the client obtains an identity assertion grant from the IdP and exchanges it for an access token at the MCP server's authorisation server, so the user never sees a per-server consent screen¹. Access follows groups and roles the IdP already knows. The announcement named Okta as the first supported identity provider, with client support from Anthropic and Visual Studio Code and a list of servers already on board¹.

My view: this is the change that makes agent access reviewable by a security team. It does not remove the need for least-privilege scopes per tool or for logging tool calls, which I cover in my MCP security checklist. But it puts the "who may use which server" decision where every other access decision already lives.

The stateless MCP spec arrives in beta

On 29 June the project shipped beta SDKs for the 2026-07-28 release candidate across its four Tier 1 languages². The headline change is a stateless core: no initialize handshake, no protocol-level session, so any server instance can handle any request behind a plain round-robin load balancer².

Two details from the post matter for anyone running servers in production. First, nothing breaks on the release date; it is when the normative text publishes, not a switch-off². Second, method and tool names now travel in HTTP headers, so gateways and rate limiters can route and meter without parsing request bodies². That is a direct win for the gateway pattern I recommend for enterprise agent platforms.

Telemetry pipelines go columnar

OTel-Arrow Phase 2 asked what happens if Apache Arrow is used not just on the wire but as the in-memory representation inside the pipeline³. The team built a Rust Dataflow Engine around OTAP, the OpenTelemetry Arrow Protocol, and compared an Arrow-native path with an OTLP-compatible one on the same runtime.

The project reports the OTAP path at roughly 10 to 20 times the throughput of the OTLP path, with 2.47 million logs per second on one core against 121 thousand³. The authors are careful to say these are measurements of specific benchmark paths, not universal claims about every Collector deployment, and I'd read them the same way. Still, for anyone whose observability bill grows faster than their traffic, the direction is clear: conversion boundaries are where the cost hides.

Two smaller OpenTelemetry items round out the month. OpenCensus compatibility requirements were deprecated in the specification, with removal no earlier than June 2027⁴. And a post from a Microsoft maintainer argued against building thin wrappers over the OTel API, showing how they quietly throw away allocation-free designs⁵. I've seen exactly that wrapper in more than one codebase.

Kubernetes: hardware and AI policy

The Kubernetes blog spotlighted the Device Management Working Group, whose main project, Dynamic Resource Allocation, had recently graduated to GA⁶. DRA is how Kubernetes allocates GPUs, network interfaces and other hardware beyond CPU and memory. If you're planning GPU clusters for AI workloads, it is now the model to design around.

The project also published how it handles AI-assisted contributions: disclose AI use, keep a human fully accountable, don't list AI as a co-author, and expect to answer review comments yourself⁷. It is a sensible template for any engineering organisation writing its own policy.

Healthcare: change control for imaging AI becomes regulation

In June the FDA published a final order classifying radiological machine-learning quantitative imaging software with a predetermined change control plan as Class II, codified at 21 CFR 892.2055⁸⁹. The special controls include labelling that the device has a PCCP, a description of implemented modifications and a version history⁸.

For architects, the significance is that planned model updates become part of the regulated design. Version history, evaluation evidence and traceability back to training data stop being good practice and start being requirements. That only works if the data underneath is governed and standardised, which is why I keep recommending a FHIR to OMOP style analytical layer for clinical AI.

What I'd do with this

1. Put agent access under your IdP. If your identity provider supports Enterprise-Managed Authorization, pilot it. If not, at least centralise MCP access through a gateway you control.
2. Test the stateless SDKs before GA. Pin exact beta versions in a staging environment and check your load balancer and gateway rules against the new headers.
3. Audit your telemetry pipeline for conversion costs. Count how many times a record changes format between source and storage. Each hop is CPU you pay for.
4. Remove OTel wrappers and OpenCensus shims from your backlog before they become urgent.
5. For regulated AI, design versioning in now. Model registry, evaluation history and data lineage should exist before the first planned update, not after.

How I can help

I help teams design agent platforms, observability pipelines and regulated AI systems that hold up in production. See my AI architecture and healthcare data architecture work, or analyse your challenge and bring the result to a call. Investors assessing AI infrastructure companies can find the same lens in my technical due diligence.