As AI models and agent infrastructure continue to develop, AI agents are able to handle more complex work over longer periods of time. This has a profound impact on product development — not only on how products are built, but on what products can do, who uses them, and how organizations must evolve to stay competitive.
Here are some of the implications worth exploring:
Embedded Intelligence
AI agents are becoming embedded in new products and services. We now talk to CRMs in natural language, marketing tools read our content and build whole campaigns, video editors can identify outtakes, and health apps nudge us toward healthier behavior. Learning Management Systems use AI to accelerate course development and personalize learning. We take autonomous vehicles for granted, and Amazon drones now drop off packages in select areas.
The products we interact with are becoming increasingly autonomous, and intelligence is now emerging as a product attribute we simply expect, just as we have historically looked for quality, ease of use, and competitive pricing.
Customization
Embedded intelligence allows us to build products and services that adapt to the needs of customers and users over time. Instead of learning about customer needs primarily through research, interviews, and observation, ongoing learning and evolution can increasingly be embedded into the product itself.
Non-Human Audiences
Since the development of the first stone axes, product development has been human-centric and ultimately driven by human needs. This is changing. Many products will have the majority of their interactions with AI agents rather than humans. This affects how we think about product marketing; we no longer need to persuade only humans. Products must be discoverable by, credible to, and purchasable by AI agents.
We now see web pages with dedicated pages for AI agents to read (see Pancake’s home page). AEO (Answer Engine Optimization) is as important as SEO (Search Engine Optimization). We predicted this phenomenon — AIs as curators of products and services — a full decade ago, and it has been interesting to watch it develop. Branding as a field must now reinvent itself, since most purchase decisions will either be AI-assisted or decided by fully autonomous AI agents.
Composability
AI agents can use protocols such as MCP, or build on-demand interface code, to connect previously independent products and services. The result is what we call Just-in-Time Assemblies (JITAs): temporary constellations of capabilities to accomplish an objective. JITAs create an almost unlimited number of use cases, which poses a challenge for product managers. What functionality should be prioritized?
Commoditization
Composability brings another difficult dilemma. If your product can be coupled with others in ways you cannot easily control, and increasingly curated by AI agents for that purpose, it is more easily replaceable. On the other hand, if you do not allow it, your product may be left behind, as humans increasingly rely on JITAs to accomplish their aims.
R&D Productivity Fuels R&D Velocity
Applying AI agents to product development can dramatically increase productivity, but that potential exists for everyone else too. The pressure to improve existing products and launch new ones will increase, resulting in a Cambrian explosion of new products and services. Speed is increasingly the most important moat, and any advantage is temporary.
R&D Velocity Fuels Organizational Transformation
In order for organizations to increase the pace of innovation, they must organize around AI-human (and increasingly AI-only) workflows. This is very different from a more conservative “transformation” fueled largely by a motivation for cost savings, where the structure of the organization does not change much. Innovation-fueled transformations are about growth and adaptation, resulting in much flatter organizations that can change dynamically.
Trust Is the Great Filter
Greater autonomy brings greater demands for trust. Individual AI agents already pose challenges for product reliability, safety, and privacy. The challenges will become greater with JITAs; your product is likely not in control of the entire value stream in which it participates.
There is an ongoing debate as to the best way to address these challenges. Emerging approaches include supervisor agents (Wayfound), run-time verification using temporal logic (Dogwood from AWS), static verification of plans using theorem provers (AWS using Lean), and neurosymbolic architectures (Imandra, Aitomatic, SRI).
In August 2026, OPAQUE Systems announced TRACE (Trust, Runtime Attestation, and Compliance Evidence), an open specification for portable, hardware-attested governance records for AI agents and other confidential workloads. TRACE was developed with AMD, Intel, and TII as founding collaborators, with support from Microsoft, and contributed to the Linux Foundation.
We can expect to see faster progress now that AI models are capable enough to pose a major cybersecurity threat worldwide. We believe that the path forward is not to curtail AI through regulation and new legal restrictions, but to solve these problems through technological progress. We will see a rich combination of static verification, attestation, monitoring, isolation, identity, certification, and other safeguards.
Context is Complicated
At the 2026 Post-Industrial Summit, which focused on Autonomous Organizations, there was tremendous enthusiasm for context engineering as a way to ensure that AI agents had the information they needed to assemble plans, use external tools, and talk with other agents. Inside organizations, this is becoming increasingly tractable.
However, when multi-agent systems cross organizational boundaries, determining what should be in a shared context becomes more difficult. Benefiting from intelligence in large multi-agent assemblies while also preserving privacy, safety, and reliability remains a difficult problem.
Compressing the Fuzzy Front End
AI agents will also transform the earliest stages of innovation. In classical product development, the so-called fuzzy front end is where we discover user needs, research market opportunities, and explore solution concepts, often by building prototypes. For physical products, decisions made during this stage will affect the cost of manufacturing, product architecture, performance, and potential quality problems.
AI agents will be able to interview potential users and customers, including other agents. They will be able to generate and evaluate (autonomously or in collaboration with humans) a large number of solutions, solicit feedback on them, run simulations, and explore solution spaces that would be much too expensive for humans to investigate exhaustively. For example, NVIDIA has developed “autonomous AI engineers” that generate, simulate, test, and optimize engineering alternatives.
The fuzzy front end will be less fuzzy, shorter in duration, and produce a vastly larger range of compelling product concepts.
Humans in Autonomous Innovation and Product Development
The initial focus for Generative AI in product development was improving task productivity. As models and AI agent harnesses improved, tasks became more elaborate. We have observed the emergence and evolution of mixed AI-human work, with humans acting as conductors of AI agent orchestras.
Speaking at our 2026 summit, Atlan CEO Prukalpa Sankar drove the point home when she explained that she did not allow engineers to develop software directly; they had to make an AI agent perform the task. This represents a bottom-up path to autonomy, and it is representative of where leading organizations are today. Humans are orchestrating AI agents.
A top-down path to autonomous product development should also be explored. Here the orchestration relationship is reversed. AI agents can manage the overall product development effort, delegating to AI agents or humans depending on fitness for purpose. This will lead to more sophisticated decision-making based on economic tradeoffs, including capability, cost, speed, and risk.
In some domains, human service work may also be embedded in the product itself, with AI agents orchestrating their work and interacting with clients. We believe this will soon be the dominant approach for providing human services, and there is a big opportunity to build the next generation of marketplaces and solution platforms for this purpose.
As AI models and agent frameworks increase in capability, we will see safe, reliable, and trusted products and services developed entirely by and for AI agents, and most of them will be invisible to humans because they will be “infrastructure” forming part of the Global AI Agent economy. This may seem unsettling to some, but consider how we are completely unaware of the vast majority of organizations that make up the economy. We don’t need to know, as we only interact with a tiny subset of them.
Where do we go from here?
Autonomous Innovation and Product Development promises to invalidate many assumptions we have had about product development, while bringing exciting new opportunities and interesting dilemmas.
We invite you to explore this topic further at our next Post-Industrial Forum event in Menlo Park on September 16, 2026, from 5pm-8:30pm
