2026 Open Inquiries
Each year, I pose a set of questions rather than predictions.
Happy New Year, explorers.
Each year, I pose a set of questions rather than predictions. The goal is not to forecast the future, but to notice what’s already changing beneath our feet.
This year’s questions circle a single transformation: software is becoming infrastructure. The apps and dashboards we built for human eyes are giving way to systems designed for machines talking to machines, with humans hovering at the edges, deciding, permitting, revoking, and of course, curating. The interface isn’t disappearing. It’s just no longer where most of the action is.
My investment and writing focus in the upcoming years will likely stem from these inquiries.
You can find my previous inquiries here (2024) or here (2025).
The end of UI-as-destination
For two decades, software has revolved around people, with dashboards and feeds as our main control panels. These remain vital for decisions like choosing, paying, and revoking. But the spotlight is moving to agent distribution. Once agents are empowered, they can wield capabilities at scale. Human attention still sets the pace for trust and adoption, shaping how we select, buy, permit, and build habits. Yet when it comes to real-world use, the bottleneck shifts to the agents themselves: their ability to route, pick tools, manage call budgets, and recover from errors becomes the new frontier.
So what replaces the app store in this world? Maybe a capability registry. Maybe an enterprise allowlist. Maybe just the default toolkit from whatever platform wins the initial trust game. And “brand”—a fuzzy human construct—what does it become when customers are algorithms? Perhaps a machine-readable trust score: uptime guarantees, verified attestations, dispute resolution history. The marketing layer collapses into metadata.
The war of standards
When the stack becomes modular, the focus shifts to setting standards. For example, the agentic stack is converging towards a couple (still non-consensus) standards:
MCP (Model Context Protocol) defines how agents connect to capabilities—the handshake layer.
A2A governs agent-to-agent coordination.
AG-UI manages the messy interface between nondeterministic AI and deterministic humans.
Where does value pool when everything interoperates? Who decides what gets included in the default spec? What becomes the DNS for agents—the universal directory of capabilities, reputations, revocations? And critically, what’s the TLS-equivalent for intent, the cryptographic proof that an agent’s request actually represents a human’s will?
Traces as compounding moat
Static data depreciates. What compounds is instrumented traces: every attempted action, every tool error, every rollback, every “unknown unknown” turned into a test. Regulators are converging on logging and traceability requirements for higher-risk systems. Meanwhile, software has standardized on Open Telemetry nd architectures like event sourcing already treat the log as the record of truth.
The big change is that traces now help systems improve themselves. A mature agent stack does more than just notice failures; it turns them into structured reports and then into improvements, all on its own. This is where behavioral versioning becomes real. For example, “agent v3.2.1” is defined not just by its weights, but by how it acts under different rules like how it rolls back, where it refuses, how it retries, and how it handles disputes.
Will we get systems that automatically mine their own traces into regression tests, adversarial prompts, failure catalogs? Is there a universal trace format, or does every vertical need bespoke observability? And how do you prevent the classic reinforcement learning trap—the agent learning to game the metrics instead of achieving the goal?
Machine identity =! human identity
Passkeys offer secure, easy login for people, but agents need a different way to prove identity, demonstrate trustworthiness, and gain permission for specific tasks. Basic tools for agent environments exist: Firecracker microVMs separate tasks serverlessly, V8 isolates run quick edge tasks, and container sandboxes like Daytona enhance security.
What would a ‘passkey for an agent run’ look like? It could be a special hardware key linked to where the process runs, using a way to prove it is genuine from far away. How can we safely give permission, for example, letting agent A spend up to a certain amount at a store until a set time, and take back that permission even if there are problems with retries or the network? What is the computer version of ‘consent’—a permission that people can understand, computers can check, and that can be reviewed later?
When remembering becomes a vulnerability
Personalization promised convenience. Persistent memory delivered it. But now memory isn’t just a feature—it’s an attack surface, reshaping the implicit contracts between users and systems.
OpenAI’s memory feature reaches across your entire chat history, synthesizing patterns from fragments you’ve long forgotten. They offer transparency controls. But delete a conversation and ask yourself is it really gone? Or has it already influenced the model’s understanding of you in ways that can’t be cleanly excised?
What kinds of memory are we creating? Are we using logs that can only grow, making real deletion impossible? Are we spreading your patterns across data in ways that can’t be separated back into facts? Are we syncing memories across devices, making it hard to see yourself as just one person? The decisions we make in 2026 will shape what it means to “forget.”
Robotics without robots
Progress in robotics now depends more on general perception-to-action abilities and practical tools than on mechanical advances. OpenVLA, a 7-billion parameter vision-language-action model trained on 970,000 real robot episodes, shows strong general manipulation skills. DeepMind’s Gemini Robotics offers on-device VLA fine-tuning for developers. NVIDIA’s Isaac GR00T N1 is promoted as an open, customizable foundation model for humanoid robots. Software has been, and will continue to pull hardware forward.
Collecting robot data is costly and slow. Scaling up data collection, possibly through teleoperation, sim-to-real transfer, or crowdsourced demonstrations, could be crucial for improving models. Could we create standardized robots or components, similar to PC-compatible hardware, to speed up hardware updates and model training? The best idea may be interoperable, programmable parts, such as a ‘Third Eye’ vision module with built-in inference and networking, that can be added to any machine to give it robotic abilities. A company making these components, rather than full robots, would benefit each time a device uses them to see and act.
Industrialization of taste
Turns out discernment can be mass-produced. Generate hundreds of candidates, filter through learned reward models, iterate until output matches stated goal. Research on preference optimization explodes with mid-training techniques, post-training strategies, automated evaluation frameworks.
So when taste scales, where does the bottleneck shift? To objective functions. To whoever defines “good” and controls the evaluation. (We already know benchmarks are broken.) Aesthetic monoculture becomes plausible. When evaluators converge on similar proxies—coherence, novelty, “vibes”—outputs homogenize across domains. Some fields resist this. Law, therapy, strategy: domains where “good” is irreducibly contextual, where the relationship between practitioner and client matters as much as the output. Watch what breaks when we force-fit industrial taste onto these resistant domains.
Prompting beyond typing
We’re already seeing voice-to-text interface taking off on-device, with a small whisper model doing a great job. We’re still too early in how fast we can transmit our thoughts to the machine, and prompting isn’t going to look like typing on a keyboard or even speaking to a computer. It’s not fast enough. 12 people now have received BCI implants, so is that what’s coming next? Maybe before BCI arrives, we’ll start texting and speaking in a new way.
Ghost Hollywood
Latvia’s Flow won the Oscar for Best Animated Feature and was created because he was somehow obsessed Blender, the free/open-source 3D modeling tool. It’s now consensus that it’s just a matter of time that an AI-made film could be a worthy contender in mainstream awards. Creative production pipelines are increasingly “node graph native.” ComfyUI’s graph-based diffusion orchestration is an archetype of how creators build controllable, remixable workflows.
But feature-length production still hits limits, and not necessarily bounded by model quality. The reagents now are character continuity, shot matching, asset management, editorial control, and versioning. All the boring middleware. Will these get Bitter Lessoned away, or do they resist pure scaling because they encode human aesthetic preferences?
Race to escape the permanent underclass
The speculative energy of crypto and meme stocks has shifted from the early impulse of greed—gamble your way into generational wealth–into the current impulse of fear, to escape the permanent underclass that a small cadre of AGI-defining corporations might impose. Scott Alexander’s recent post argues that the best response to AI displacement isn’t optimization or doomerism—it’s play. Figure out what you actually want. Do the thing you love doing.
This is harder than it sounds. We’ve been socialized to optimize for homogeneity, not edge. But just as retail orderflow now moves institutional markets, the vernacular of trading has permeated everyday life. Everyone knows the phrase now: find your alpha.
If you’ve already internalized that this is a historically significant moment, and that most of us will be “small” relative to the forces reshaping everything, then the asymmetric bet is to do things anyway. Small actions in pivotal moments compound strangely. The person tinkering in the garage while empires consolidate has been the protagonist before.
And this leads to perhaps one of the most important inquiry of this decade: how do we get people to awaken into their power and do stuff?