The Exokernel and Its Implications for the Future

November 2026

Insofar as they are economic problems at all, the world’s problems in this generation and the next are problems of scarcity, not of intolerable abundance. The bogeyman of automation consumes worrying capacity that should be saved for real problems …

— Herbert Simon

The operating systems of computers have one fundamental job: to manage the interface between the user and their device’s hardware. Operating systems do this with policies, and these policies entail tradeoffs. We can explore this by taking the processor on a computer. Usually there are many processes to run, so each process gets a slice of time depending on various parameters like the program’s priority, its CPU usage, and how long it has been waiting to execute. These seemingly innocuous parameters can have substantial consequences on performance, manifested concretely as user satisfaction with the device. For example, suppose certain fixed parameters limit each program to a short interval of runtime. This system is suitable for applications like video games, where the latency between button presses must be minimal (and failing to do so is the difference between using and not using the video game), but consider downloading a movie. As opposed to the quick short-lived processes, we have one long continuous process; constantly switching to other processes would elongate the download, and if we are crunched for time, say going on an airplane, this could be the difference between enjoying and not enjoying the plane ride.

By creating policies optimized for certain use cases, we are implicitly creating suboptimal policies for all other use-cases. Operating system designers handled this by finding a suitable middle ground, but we have the same problem because policies for the average use-case imply suboptimal policies for all other use-cases. Graphically, taking the y-axis as performance gain and the x-axis as different jobs, we have moved from a normal to a uniform distribution, and the mass under the curve, meaning the performance gained over all applications, remains the same.

Out of this tension came the exokernel (Engler et al.) in 1995.1 The exokernel allowed application developers to create policies optimized for their own application, and these policies could coexist among other application policies. Using the graphical example previously, the mass under the uniform distribution could increase, reaching optimal performance for all applications as opposed to one. But this entailed tradeoffs. In particular, more work for the application developer. They would now need to create policies tailored for their application’s use case, and, truth be told, they did not want to do this. What’s more, by then the incumbent Unix and Windows operating systems were too entrenched to replace; switching to this design was not worth it. Overall, some of its ideas filled narrow corridors of operating system design, most notably with the extended Berkeley Packet Filter mechanism in Linux, but the overarching idea is a relic in operating system history.

But relics, if for nothing more than serving as the means for intellectual inquiry (as this essay is), can still have utility. Philip Wadler, a computer scientist known for his contributions to type theory, noted in a talk at Strange Loop that it takes roughly 20-30 years for a great research idea to become applied. The example he used was linear types, which, in the closely related form of affine types, became a hallmark of the Rust programming language, providing programmers with the ability to reason about memory usage algebraically and therefore algorithmically, endowing the language with the performance guarantees of C without manual memory management and the security vulnerabilities that ensue, no small feat within computer science. Could something similar happen to the exokernel?

The exokernel design captures the idea of tailoring a system’s parameters to a specific user and their applications, in contrast to the monolithic kernels of today that provide a fixed set of system parameters for all applications. Generalizing this mechanism, we find a similar principle separates subjectivity and objectivity in the sciences. Objectivity describes the scientific method as largely void of the scientist’s judgement, using established rules, methods, and principles as opposed to their expertise. Newtonian physics is the canonical example; the same principles can be observed regardless of the user’s intent. Subjectivity prioritizes the perspective of the scientist and their own experience, with psychoanalysis as the best example. Therefore, in a more subjective science the “parameters of the scientist’s psyche” are emphasized for their work. Might we be seeing a resurgence in subjectivity, and if this is the case could the exokernel be the correct architecture, now at the level of the user?

We can observe this shift to subjectivity today in computer science interviews. With the introduction of coding agents there no longer exist objective measurements of a computer scientist’s ability (i.e., does the code pass the tests or not?). Instead companies are following Palantir’s model, throwing the interviewee into some blank slate problem they have to work through. Their chain of thought is critical here, similar to the psychoanalyst. But with this introduction of coding agents another shift has been happening. In particular the introduction of hobbyists. During the 1950s to the 1970s hardware was in the middle of a large transition phase. It had moved from large clunky components like vacuum tubes to elegant transistors, and resided in a unique middle ground where it was complex enough to be exciting, but simple enough to understand. So building and hacking flourished, with audio equipment, radios, TVs, clocks, alarms, and more among the most common creations. Today, this corresponds with the introduction of coding agents and simple tools like Python and Amazon Web Services. Computer scientists are no longer limited by their understanding of the tools themselves. That is, the bottleneck shifted from the training required to use the tools, to the scientists’ own ideas, values, and chain-of-thought. Hobbyists are subjective scientists in this aspect, and the exokernel could provide the optimal design for a “knowledge work machine” for these subjective scientists.

I believe the present AI-consulting companies hint at this, but the types who create these companies do so for the means of exploitation and are not necessarily visionaries. Paul Graham, with respect to AI and the corresponding academic output in the 1980s, remarked: “Its brokenness did, as so often happens, generate a lot of opportunities to write papers about various band-aids that could be applied to it, but it was never going to get us Mike.”2 The current forward-deployed engineers and AI consulting companies like Tenex and Flowscope (the latter is ironically funded by Y-Combinator) apply patchwork solutions to the current software used by companies, but they fix a system that is fundamentally broken, which gives the illusion of a successful business model.

To drive this point home imagine life before the personal computer. A company may have had the need for various technologies. A forward-deployed engineer might come in and notice the need for a timer to indicate when a meeting ends, a clock in every office room for indicating the time of day to keep employees on track, a telephone or intercom for communicating with employees in different rooms, or a distributed system of calendars for managing and coordinating different employee schedules. They will develop a separate system for each need, but these bandaids will miss the crucial insight that a single device could encompass all of these functions: the desktop computer. The hobbyists who tinkered with hardware were the ones able to collaborate and produce the technology that allowed for all of this to be incorporated into this abstract machine.

After all this did end up being the general flavor of the personal computer (which is why the exokernel was such a promising design). Every desktop looks different because of our own unique applications. I have Serato and Ableton because I DJ (I am 23 and went to college in the northeast so this should not come as a surprise), but I also have VSCode and Cursor because I am a computer scientist by trade. Someone else might have iMovie and Figma because they do video-editing on-the-side and Excel because they are an analyst at a financial firm. The point is our subjective psychology induces the need for certain tools and applications. The exokernel is uniquely suited for providing this and may be the required model for this next machine. That is, a new and novel piece that can reflect the user’s own values and intelligence, and this piece may even evolve with the user. It is fun to imagine what this could be. It might be software that is able to communicate telepathically by means of EEG headsets, or by analyzing the slight unconscious contours of our face at any given moment. Or maybe it is software that can change its UI/UX autonomously according to various latent factors. In a way, the “Second Brain” formulation might be this revolutionary technology, though certainly not in the capacity we see it today.

The use of AI

I wrote this entirely myself. I asked AI questions, usually paraphrasing boring facts but never ideas, though the facts AI bore did inspire some.

Notes

  1. Dawson R. Engler, M. Frans Kaashoek, and James O’Toole Jr., “Exokernel: An Operating System Architecture for Application-Level Resource Management,” SOSP 1995. https://www.cs.utexas.edu/~witchel/380L/papers/engler95sosp-exokernel.pdf ↩

  2. Paul Graham, “What I Worked On,” 2021, https://paulgraham.com/worked.html. “Mike” is the self-aware computer in Robert Heinlein’s The Moon Is a Harsh Mistress (1966). ↩