The Agentic Infrastructure Stack: MCP, ARD, and the Standards Building Production AI Agents in 2026

Infrastructure tends to get less attention than models, but infrastructure is what determines whether a technology ships to production or stays in the lab. For agentic AI in 2026, the infrastructure story is suddenly the most interesting one. Model Context Protocol, Agentic Resource Discovery, background execution APIs, and new process-level benchmarks are converging into a coherent stack — and enterprises are noticing.

Here is what is actually being built right now, who is building it, and why it matters for anyone shipping agents.

The Discovery Problem Is Now a Protocol

Every agent builder eventually hits the same wall: the agent needs a tool, and the user is expected to know which one, where it lives, and how to configure it. Hardcoding MCP server URLs into config files does not scale. Dumping every available tool description into the LLM context window hits budget limits fast.

The Agentic Resource Discovery (ARD) specification is the answer, and it is moving fast. Developed as a draft open standard with contributors from Microsoft, Google, GoDaddy, and Hugging Face, ARD defines how capabilities are cataloged, indexed, and searched across federated registries so agents can find what they need at runtime.

ARD is deliberately not a product or a marketplace. It is a shared envelope. Publishers host capabilities at a well-known URL using a static ai-catalog.json manifest. A registry exposes a POST /search endpoint that ranks results by publisher identity, representative queries, compliance attestations, and tags. The client searches in natural language, and the model invokes whatever comes back.

Hugging Face shipped the first reference implementation, Discover, which wraps the Hub’s semantic search and serves results as ARD catalog entries. An agent searching for “fine tune a sentence transformer” can receive back Skills, MCP servers, or raw Space metadata depending on what the client asks for. Because federation is built in, a search through one registry can surface capabilities hosted by another.

The shift is meaningful: from install-first, use-later to intent-based, runtime discovery. That is the difference between a plugin ecosystem and an actual agent ecosystem.

Google’s Managed Agents Grow Up

On July 7, Google expanded Managed Agents in the Gemini API with four capabilities that directly address production pain points. Background execution lets agents run asynchronously on the server, returning an ID for polling instead of requiring a fragile long-held HTTP connection. Remote MCP server integration eliminates custom proxy middleware by connecting managed agents directly to external Model Context Protocol endpoints. Custom function calling lets developers mix built-in sandbox tools with their own local business logic, with step matching routing each tool to the right place. And credential refresh means network configurations can be updated across interactions without losing sandbox state.

These are not features for demos. These are features for systems that run for hours, touch internal databases, and handle real credentials. Google is explicitly positioning Managed Agents as asynchronous workers inside real development environments — not sandboxed toys.

OpenAI Ships the Silicon, Not Just the Software

While Google built the API layer, OpenAI went deeper into the stack. In late June, OpenAI and Broadcom unveiled Jalapeño, a custom AI chip built specifically for LLM inference. The chip is designed to improve performance, efficiency, and scale across AI systems — and it signals that the entire hardware layer is being retooled for agentic workloads.

OpenAI also previewed GPT-5.6 Sol, pairing stronger coding, science, and cybersecurity capabilities with its most advanced safety stack. The model and the silicon are being co-designed.

Meanwhile, OpenAI’s internal data showed just how quickly agents are becoming the default interface. By April 2026, every department at OpenAI — Legal, Finance, Recruiting — had crossed into agents as their primary AI tool. Non-developer Codex users multiplied 137 times among individual users and 189 times inside organizations. At the 99th percentile, daily active users generate more than 60 hours of agentic work per day across parallel agents.

Benchmarking the Process, Not Just the Answer

A quiet revolution in evaluation is underway. Hugging Face published two agent-focused benchmarks that measure what actually matters in production.

ScarfBench evaluates AI agents on enterprise Java framework migration — exactly the kind of tedious, high-stakes codebase migration enterprises pay consultants to perform. It matters because enterprise adoption depends on whether agents can handle legacy systems, not just greenfield projects.

The “Is it agentic enough?” harness goes further: it does not ask whether an agent gets the right answer. It asks how much work it takes to get there. Turns, tokens, seconds, whether the agent walked a clean path or used deprecated APIs. Two agents can both return the same classification result, but one writes a 40-line script with three re-runs while the other types a single CLI command. Both succeed. One is dramatically more expensive.

Without process-level benchmarks, organizations optimize for accuracy while bleeding tokens on inefficiency. That gap is now being closed.

Security and Science Move In

As agents gain autonomy, the security conversation is maturing. Anthropic redeployed Fable 5 globally on July 1 and proposed an industry-wide framework for scoring jailbreak severity alongside Amazon, Microsoft, Google, and other Glasswing partners. Cohere introduced North, a security-focused AI agent built in partnership with Wiz. And Hugging Face’s MosaicLeaks benchmark poses a direct question: when a research agent has access to sensitive data, can it keep a secret?

Anthropic also shipped Claude Science, an AI workbench for researchers that integrates scientific tools, produces auditable artifacts, and provides flexible compute access. It is a signal that agentic interfaces are expanding beyond engineering into domains where reproducibility and provenance matter as much as speed.

The Open Stack Fills the Gaps

Closed platforms do not own this story. OpenClaw’s July beta added GPT-5.6 support, external harness attachment for interactive Codex workflows, and event-driven cron scheduling. Hugging Face demonstrated local models triaging OpenClaw pull requests at no cost and made it possible to run a vLLM server on HF Jobs in a single command. Mistral’s Vibe agent launched Work and Code modes with remote agent capabilities and MCP connectors. Le Chat gained custom MCP connectors and persistent memory.

The pattern is consistent across projects: agent-optimized CLIs, Skills for discoverability, self-contained examples, and MCP-based tool integration. The open-source ecosystem is converging on the same architectural primitives as the closed platforms.

Enterprise Deployment Accelerates

Behind the tooling headlines, buyer behavior is shifting. Recent survey data from 830 IT leaders shows agentic AI ranked as the #1 enterprise technology priority. Enterprises are scaling agentic AI at more than double the rate of mid-market companies. Nearly 39% of buyers now expect generative AI to be delivered primarily through autonomous agents.

The pilot phase is ending. The deployment phase is beginning. And the organizations that win this phase will be the ones that invested in the infrastructure — discovery, security, observability, and benchmarking — while others were still optimizing their prompts.