Walk into any AI infrastructure meeting these days and you hear the same conversation. Different rooms, different speakers, different titles on the badges. The same words. Agent. Identity. Trust. Governance. Standards. Everyone nodding. Almost nobody asking what those words mean in practice. The agreement is real on the surface and thin underneath, and the underlying technical understanding ranges from genuinely deep to surprisingly shallow to actively confused. The shape of the conversation is consistent. The depth of the conversation varies wildly.

This is the state of the AI industry one year into the agent moment. The hype is at an all-time high. The investment is at an all-time high. The number of vendors selling products with “agentic” in the name is at an all-time high. Underneath the surface, the actual practical knowledge of what we are building, why, and what good looks like is significantly thinner than the surface activity suggests. And in the middle of all of this, the industry is in a mad rush to write standards for technology it has yet to fully understand.

It is the Wild West out there. The conditions that produced the Wild West remain in place. The Wild West is going to continue being the Wild West for at least another year, possibly two. The question is whether we use this period to lay durable foundations or whether we lock in choices we will spend the second half of the decade unwinding.

The hype is real, and most of it is justified

The first honest observation is that the hype is mostly downstream of something genuine. Language models can do things that were science fiction four years ago. Reasoning has improved dramatically. Tool use has become reliable enough to deploy. Agentic systems are doing real work in real companies, and the productivity gains in narrow domains are large enough to justify the investment.

A lot of the noise is signal. Companies are spending real money because real value is appearing. The pace of capability improvement remains aggressive. The macro story is sound.

But the macro story being sound is no excuse for the micro story being sloppy. Inside the hype envelope, the quality of execution varies enormously. Some companies are doing serious agent infrastructure work with thoughtful architectural choices. Many companies are shipping prompt-chained automation and calling it agentic. Most of the agent identity products on the market are rebranded service principals with a marketing budget. A surprising fraction of “AI governance” deployments are application-layer logging with a compliance dashboard bolted on.

This gap between the hype and the substance is itself a feature of the moment. It is what every technology era has looked like in its first eighteen months. The web in 1995 was mostly hand-coded HTML in animated GIFs. The mobile era in 2008 was mostly websites with bigger buttons. The cloud era in 2012 was mostly virtual machines with new branding. The first eighteen months of every era is mostly people doing the old thing with the new label, and the patterns that will matter for the next decade are still being discovered.

The risk is when we mistake the first eighteen months for the destination, and we standardize around it.

The knowledge is thin in specific places

The shallowness of practical knowledge in this moment shows up unevenly. Some areas are genuinely deep.

The ML research community has spent years building the foundations the current moment rests on. Model architecture, training methods, evaluation frameworks, alignment work. The depth here is real, and the people doing the work are mostly clear about what is known and what is being figured out.

Application development has matured fast. Engineers building agentic features inside products have learned a lot in the past year about prompt structure, tool integration, evaluation harnesses, failure modes. The hands-on knowledge inside engineering teams that ship working systems has compounded faster than expected.

But the infrastructure layer between models and applications, the layer where standards are being written, is where the knowledge is thinnest. The people writing identity specs for agents have mostly never deployed an agent at scale. The people designing governance frameworks for agents have mostly never been on the receiving end of an actual incident involving one. The people standardizing agent transport have mostly never operated a multi-organization agent network in production.

This is normal at this stage. The infrastructure layer always lags the application layer in maturity, because the infrastructure layer crystallizes what the application layer has learned. But the infrastructure layer is also being asked, right now, to publish standards that will shape the next ten years of agent deployment. And the gap between the standards being published and the practical experience they would normally codify is wider than usual.

The standards rush is doing things in the wrong order

Standards historically work best when they ossify hard-won practical knowledge. HTTP/1.0 was published in 1996, six years after the web started. SMTP was standardized in 1982, after a decade of operational email systems. DNS came after the HOSTS.TXT model had run aground at scale. TLS came after SSL had gone through three iterations and several real-world attacks. OAuth 2.0 came after years of bilateral integration patterns had shown what worked and what failed.

In each case, the standard was a synthesis of accumulated experience. The community had implemented the thing many times, watched what broke, learned what mattered, and then encoded the lessons. The standard was good because the experience was deep.

The agent standards being written today are mostly happening in a different order. The specifications precede the deep operational experience. Working groups are publishing drafts for components that have yet to see large-scale production stress. Partner rosters are announced for specifications that have yet to ship reference implementations. Industry analyst firms are publishing market maps for categories that are eighteen months old.

None of this is a complaint about the people doing the work. Most of them are doing their best with limited information, and somebody has to start writing these specifications. The alternative (wait for complete consensus before publishing anything) produces the vendor-fragmented landscape that gave us shadow IT and credential leak inventories. Action under uncertainty is the only available option.

The risk is that the specifications being written in 2026 will get locked in as the foundation for the next decade, and the operational experience that should have shaped them will arrive too late to influence them. The standards rush is producing artifacts the field will have to live with for a long time, written from a knowledge base that is shallower than the artifacts pretend.

What I have learned watching from inside

Spending a year inside the standards rush as a participant has produced some observations I lacked when I started.

The most striking is that the people most committed to writing standards quickly are often the people with the most to gain from a particular outcome. Vendors with existing products want standards that look like their products. Working groups with existing primitives want standards that extend those primitives. Companies with existing customer bases want standards that protect those bases. The participants who would benefit most from waiting for deeper knowledge are often the participants with the smallest seat at the table.

The second observation is that the structural choices being made now are larger than they appear in the moment. A decision about whether identity is application-layer or transport-layer changes what the next decade of products can look like. A decision about whether discovery is centralized or federated changes who controls the market. A decision about whether attribution is signed or logged changes what regulators can actually do. These choices feel like spec details. They are infrastructure for an industry.

The third observation is that the writing-in-public model the IETF and similar bodies use is doing better than it gets credit for. The drafts are open. The criticism is open. The revision history is open. The mistakes get caught and fixed in ways that closed standards processes would never permit. Watching the working group dynamics from inside has made me more confident in the process, even when I disagree with specific outputs. The mechanism is healthier than any individual draft.

The fourth observation is that the participants who are doing the best work are the ones who have implemented something at production scale. They know what the protocol needs to carry because they have watched it fail to carry the right things. They know which primitives matter because they have watched applications break when those primitives were absent. The gap between participants who have shipped working systems and participants who have only written about them is wider than it looks from the outside, and it shows up in every meeting.

What the next year requires

If the standards rush is producing artifacts the industry will live with for a decade, the next year is when the artifacts have to get good enough that the decade is survivable. Some things would help.

Implementation before publication. Working groups should require reference implementations that have been deployed at non-trivial scale before specifications are advanced to final status. The IETF has a “running code” tradition. The agent space needs to honor it.

Revision-friendly architecture. The specifications being published now should be designed for revision. Versioning, deprecation, federation across versions, graceful degradation. We are going to learn things in 2027 and 2028 that we cannot anticipate today. The standards should be ready for that.

Honest acknowledgment of what remains unknown. Specifications should name the open questions, rather than pretending they are settled. Working group drafts should include sections on what remains uncertain. The current culture of confident specifications papers over the genuine uncertainty in a way that hurts the quality of the result.

Wider seats at the table. The participants most affected by standards (regulated industries, public-sector deployers, civil society) are underrepresented in the standards bodies writing them. The participation cost of working group engagement is too high for organizations that have the most to lose from bad outcomes. The processes need to find ways to lower that cost.

Patience where patience is possible. Some specifications can wait. Some categories of infrastructure can run on ad-hoc patterns for another year while the practical knowledge catches up. The rush to standardize everything at once is making some choices worse than they need to be.

The closing observation

The AI industry in 2026 is real. The hype is justified by genuine capability and genuine value. The investment is going to specific things that work. The standards rush is happening because somebody has to make the calls, and the people making them are mostly doing their best.

It is also still the Wild West. The depth of practical knowledge is uneven. The standards being written are running ahead of the experience that should have shaped them. The choices being locked in now will shape the next decade. And the participants with the most to gain from particular outcomes are often the ones most aggressively pushing those outcomes.

This is the moment we are in. The work that matters right now is the work of making good choices under genuine uncertainty, with humility about the limits of current knowledge, and with enough discipline to revise when we learn better. The hype will keep climbing. The standards will keep getting written. The knowledge will keep catching up.

What the next year requires is a different relationship to the moment than most of the industry is currently having. Less certainty in the specifications. More acknowledgment of what remains uncertain. More implementation before publication. More humility in the working groups, and more skepticism of any standard that arrived without practical stress-testing.

The Wild West will end. Some of what gets built in it will become permanent infrastructure. Some of it will be torn down later. The work right now is being awake enough to tell the difference, and to choose carefully which category we are contributing to.

That is the report from the inside. The view from outside, I suspect, looks even more chaotic. The view from five years out will probably look like a foundation being laid in a hurry, by people who knew less than they pretended to, and which somehow still got mostly right. That is how every era ends up looking in retrospect. The work is to make this one end up that way too.

---

If you find this content valuable, please share it with your network.

Follow me for daily insights.

Book me to speak at your next event.

Start managing your agents for free.

Chris Hood is an AI strategist and author of the #1 Amazon Best Seller Infailible and Customer Transformation, and has been recognized as one of the Top 30 Global Gurus for Customer Experience. His latest book, Unmapping Customer Journeys, is available now!