A Few Things We’re Watching in 2026

Each year, we take stock of the shifts reshaping how companies are built and where founders might uncover unfair advantage. If 2025 was the year AI became truly operational, 2026 is the year it becomes fully embedded: in messy workflows, complex org structures, labs, policy, design cycles, and the unwieldy backends of how businesses actually run. 

If there’s a through-line across these predictions, it’s that AI is moving closer to the real work. The founders who win in 2026 won’t be the ones chasing abstraction; they’ll be the ones who know exactly where to land inside the stack and how to build something that makes a real impact.

Here’s what’s rising to the top of our radar for the year ahead.

_{PREDICTION#01}

Computer-Using Agents Become the Next Infrastructure Unlock

The last wave of AI brought new input modalities, namely voice, chat, and vision, and with them, new distribution wedges. In 2026, computer-using agents mark the next leap. These systems don’t ask for API access; they operate software the way humans do.

But the wedge won’t be generic “agent copilots.” It will come from startups that pick a specific workflow and go deep, collecting contextual artifacts (instructions, recordings, UI states), rethinking the process, and automating an entire job function end-to-end. The biggest gains will start in the back office, where labor spend is highest and tools are oldest, before expanding outward as reliability grows.

_{PREDICTION#02}

The Wall Between Design and Engineering Finally Falls

For years, product development has been defined by handoffs: PM → design → frontend → backend. In 2026, AI-native creation stacks collapse the chain.

Designers will express intent and receive runnable, testable software…no translation layer required. The result is faster cycles, smaller teams, and a new company archetype built around the “full-stack designer.” This may become one of the most culturally disruptive shifts in product orgs in decades.

_{PREDICTION#03}

Prediction Markets Explode, Paving the Way for New Fintech Opportunities

After clearing regulatory hurdles, prediction markets have gone mainstream and are still only in the early innings of adoption. In the coming years, we will live in a society where anyone can wager on any event at any time. Love it or hate it, Kalshi and Polymarket have opened the floodgates for new companies to support and enhance this ecosystem. What can we do with the insights gleaned from more-perfect information on events? Will we see insurance products that hedge and protect ourselves from these new risks? Will a new crop of hedge funds emerge to find alpha in event-based trades? We’ll see founders explore these areas in 2026.

_{PREDICTION#04}

Biology Gets Rewritten for Next-Generation Therapeutics

AI is transforming discovery, but drug development and other medical advances remain tethered to biology’s slow feedback loops. 

Traditional labs, regulatory hurdles, and high-risk experiments make AI-only solutions hard to value and adopt. Improvements in AI-guided RNA and molecular design, coupled with faster experimental automation and regulatory alignment, will let platforms deliver measurable therapeutic outputs at scale - not just in preclinical experiments, but throughout clinical trials, shortening timelines and improving candidate success rates. Suddenly, one AI model can power dozens of programs, dramatically increasing ROI per dollar invested.

There is room for a platform-first biotech to fundamentally shift how drugs are discovered, developed, validated, and brought to patients.

_{PREDICTION#05}

AI Develops Physical World Capabilities

AI is mastering perception, but the physical world remains stubbornly difficult. 

Robotics and manufacturing still rely heavily on human intuition, teleoperation, and labor-intensive setup, meaning most AI-enabled robots today tackle narrow, highly curated tasks. Even with cheap compute, low-cost sensors, powerful ML, and abundant cameras/lidars, scaling general-purpose robots has been bottlenecked by data collection, environment adaptation, and task diversity.

We expect a tipping point: sample-efficient learning, modular tele-op co-training, and scalable foundation models for robotics, combined with vertical and task-specific post-training, will let machines learn new tasks from less than an hour of demonstration data and generalize across environments. Once robots can autonomously execute a broad range of real-world tasks, from warehouse picking to folding, assembly, and cabling, without weeks of custom setup, AI will finally unlock physical work at scale.

_{PREDICTION#06}

Payers Finally Crack the Code on AI Reimbursement

AI is reshaping healthcare but its impact has been mainly focused on the back office. Yet the crux of the American healthcare burden is in care delivery. Reimbursement formulas have stalled progress as CMS (the Center for Medicare and Medicaid Services) currently values human labor as a necessary component of cost, which means AI-only services are priced at zero.

We expect a policy shift. Once CMS updates its methodology to reward efficiency rather than human input, the floodgates will open: new business models, new incentives, and far more room for automation in one of the largest markets in the U.S. This is the blocker. Removing it unleashes an entirely new landscape.

_{PREDICTION#07}

Open Source Closes the Gap… and Expands the Frontier

Silicon Valley is increasingly powered by open-source models, many coming from China, which are not only viable alternatives to closed labs but often outperform them, especially in multimodal tasks.

As long as Chinese tech giants continue releasing strong open source work, we’ll see more startups choose these models for cost, performance, and speed. The big question: can American players like Meta or Reflection catch up? And can the RL-as-a-service ecosystem prove that open models can truly stand on their own? 2026 will be a decisive year.

_{PREDICTION#08}

Computer-Use Breakthroughs Make the Forward-Deployed Engineer Obsolete

Despite the explosion of new AI applications, enterprise adoption is still slowed by something surprisingly analog: the need for people to configure, deploy, and train every new system. Post-sales onboarding, user training, and forward-deployed engineering all scale linearly with revenue and often stretch implementation cycles to a year or more. It’s a strange mismatch for a sector defined by automation: we have AI powerful enough to reason and act, yet we still rely on humans to walk customers through setup on Zoom.

Advances in computer-use agents will change that. Because these agents can operate software the way people do, like navigating UIs, configuring backend integrations, triggering API calls, etc., they can take over large portions of post-sales and deployment work. They can stand up environments, guide users through live training sessions, and troubleshoot in real time without a human in the loop.

This shift collapses go-live timelines, reduces the need for massive services teams, and finally lets AI applications scale at the speed the technology promises, not at the speed companies can hire.

‍_{PREDICTION#09}

AI Infrastructure Rebuilds Itself Around Safety, Control, and Real Work

2026 is the year AI steps out of the realm of “chat and code” and into the operational core of the enterprise. Agents will finally get access to real work products - live systems, sensitive data, and actions that actually matter. They’ll run code that hasn’t been pre-checked by a human, operate in production environments, and touch workflows where mistakes are costly.

That shift forces a fundamental rethink of enterprise infrastructure. Speed alone won’t cut it. To let agents work inside the business, we need stronger guardrails, granular permissions, real-time observability, and circuit breakers designed for a world where software can act autonomously.

The cloud era was built on trust in the human operator. We assumed employees wouldn’t intentionally wreck their own systems. Agents don’t get that benefit of the doubt. They require infrastructure that assumes nothing, monitors everything, and limits the blast radius by design.

The companies that win this next phase won’t just make AI faster; they’ll make it safe enough to matter.

_{PREDICTION#10}

Consumer AI Finally Becomes Personalized

AI is supposed to know us. At least that’s been the promise dangled tantalizingly over us for the past several AI hype cycles. However, the products we get are still broadly generic and generalizable - trying to be all things to many people. In part, this is due to the limitations of AI memory, which has been an impediment to making real progress in building consumer AIs that feel real and more human-like. Still, we are on the edge of several breakthroughs in agentic memory that will unlock a myriad of real-world use cases ultimately allowing for much more continuity in our AI interactions. AIs will be able to remember things about us, and take actions behind the scenes based on our past musings, preferences, and ideas. This new foundation will open up a lot of much more magical user experiences and - faced with such obvious utility - cause many of the AI skeptics to rethink their points of view.