· ai-devtools
DreamCoder, AI coding harness and IDE, released by author of Mastering Emacs
After fifteen months of tireless refinement, a major project is finally ready for formal release, marking a welcome break from a period of relative quiet.
This new development, DreamCoder, is an AI coding harness and IDE that diverges significantly from existing tools, and the author intends to discuss it alongside a brief history of AI, LISP, and Emacs.
The author recently put Claude Code to work, installing it shortly after its debut in May 2025. By tasking it with building a complicated piece of enterprise middleware—a project the author had tackled numerous times throughout their career—they witnessed an astonishing performance. Observing the tool process the project with that characteristic, early Sonnet energy was remarkable. Although the output was far from flawless, demanding considerable human guidance, input, and correction, the sheer speed advantage it offered was undeniable. Furthermore, watching it transition between various programming languages, frameworks, and concepts, moving fluidly from theory to practice in each turn, was a truly captivating spectacle.
One aspect of Claude Code that caught the author's attention was its reliance on a terminal application, something they hadn't used for such an agent in a long time. This felt like an unusual choice for a coding assistant, leading to a trend where other agents followed suit, adopting terminal interfaces. The author questioned the necessity of staring at a terminal window that seemed as large as a shoe rack to perform coding tasks, noting that while GUI versions exist, the terminal interface remains the dominant mode for most such utilities. Moreover, the author observed that most current harnesses, and most GUIs, treat graphical interfaces as secondary features, often presenting a monotonous design featuring a side panel for chat headings and a corresponding chat area. However, a dedicated programmer requires substantially more than this basic setup; even a casual user finds the process of copying and pasting multi-line text in a terminal cumbersome. Consequently, in May 2025, the author resolved to construct a superior harness, believing that with modern AI, this endeavor should take no more than two or three months. Following 15 months of persistent iteration across both the user interface and the backend, and after successfully completing a beta test, DreamCoder is now ready, representing, in the author's view, the pinnacle of what a coding harness ought to be.
Before delving into the specifics of this creation, it is worthwhile to explore the historical context of AI and Emacs's role within it. Tracing back to the 1950s, both Emacs and LISP hold a distinguished place in the narrative of artificial intelligence. John McCarthy, an AI researcher at MIT, initiated the creation of the Advice Taker during that decade. This software aimed to encapsulate a framework for what he termed "commonsense reasoning," utilizing "mathematical logic." His core hypothesis suggested that if we could express every facet of humanity—such as common life facts—with symbolic precision, we could initiate the fundamentals of AI. Back in the fifties, computer software capable of this was scarce. Formalizing such a system necessitates substantial, rigorously structured data alongside a method for manipulating that data; indeed, the data and the code itself could become recursively intertwined, rendering a standard programming language insufficient.
To aid in this effort, McCarthy developed the LISt Processing (LISP) language, which contained two components: M-expressions, functioning as a functional language akin to Algol, and S-expressions, designed solely for representing symbolic data, never intended for code writing. S-expressions were visually augmented with parentheses, while M-expressions employed square brackets to visually delineate the two structures. Uniquely, the language specification included an `eval` capability, theoretically proving the language as a universal Turing machine—had McCarthy possessed the necessary resources. This capability remained theoretical because McCarthy lacked the necessary staff and funding to build the recursive evaluator capable of processing its own world state alongside incoming expressions, a limitation made starkly apparent by the complexity of his M-expression language and the limited resources available at the time.
Steve Russell, who collaborated with him, demonstrated the limits of McCarthy's approach by succeeding where others failed: simply by executing the work. Russell managed to write the first LISP Interpreter featuring a functional `eval`, but the M-expressions were subsequently marginalized; the imperative, procedural nature of programming was instead absorbed into S-expressions. The initial process required manually scripting the language on paper, punching it onto cards, and then running it in batches. The decision to discard M-expressions was a matter of sheer practicality. Thus, LISP came into being, and through several historical coincidences, M-expressions failed to gain widespread acceptance. While M-expressions persist in Wolfram’s Mathematica, their presence elsewhere is rare. Later, McCarthy acknowledged that, “One can even conjecture that LISP owes its survival specifically to the fact that its programs are lists, which everyone, including me, has regarded as a disadvantage.” S-expressions were always meant to be the storage format, bridged by a language for manipulation. However, by embedding evaluation within the same symbolic representation as the data, we arrived at what is known as LISP today, which accounts for its unique nature.
Consider a simple calculation: `(+ 2 2)`. This structure can simultaneously represent an expression awaiting computation, such as "two plus two," or it could function as a ledger documenting a calculation already completed. McCarthy never realized the Advice Taker in its purest form; he and his team encountered numerous philosophical, mathematical, and computational barriers that occupied them for a considerable duration. The fundamental premise of the Advice Taker was straightforward: one could symbolically describe the journey to a destination in a car by constructing a sequence of narrative facts—akin to a flowchart, yet composed of small rules and facts rather than one monolithic design. The system would automatically resolve any ambiguities, requiring the logician only to capture the essential elements of the world they wished to model, such as: `(at me home) (walkable home car) (walkable home driveway) (drivable car shop)`, and so on.
Computer Science students will recognize parallels to Prolog in this concept. The original goal was simple: if we can symbolically represent the world around us, we can guide that symbolic world using real-world data, allowing the automaton using this formalized symbolic world to evolve accordingly. Yet, the inherent axiomatic complexity of modeling a world proved too daunting for any system, even one intended for a child. Every axiom introduced generates new ones—if you can walk to your front door, what about if the door is locked? Or if your vehicle is parked on the street instead of in your driveway? A rules-based depiction of reality is beautiful and delicate, but incredibly difficult to construct; it becomes a fractal of increasingly specific and pedantic rules that undermine the perceived completeness of the world model. Anyone familiar with Prolog understands this struggle intimately!
The author has a personal anecdote related to this theme. During a summer internship in the nineties, while still a teenager, at a firm developing banking software, the company seemingly had no designated role for the young interns. On the first day, they gathered the interns in a room, divided them into teams, and assigned them the task of creating a flowchart for assembling a bicycle, with the best design winning. They were instructed to be as explicit and detailed as possible, accounting for every possible condition, and the assisting employees did their utmost to help. The author distinctly recalls losing because their flowchart neglected to include the step of opening the box containing the disassembled bike.
Meanwhile, McCarthy eventually relocated and established his own AI laboratory at Stanford, though the concept of the Advice Taker persisted. In the late 1970s, Stallman and Steele developed EMACS (Editing MACroS) at MIT’s AI Lab, building upon TECO, a notoriously opaque text editor. TECO did not automatically refresh the screen; users had to manually enable this feature and hope their edits functioned correctly. Users typed brief keyboard commands, held their breath, and waited patiently for the text manipulation to take effect, repeating this cycle until the desired result was achieved. Stallman and Steele desired more functionality: they sought macros to automate the repetitive nature of typing such constant stream of keystrokes. This thread of augmentation, and later self-modification, remains a central theme in Emacs today.
Although Emacs has a long history, its ultimate trajectory is well-known. The author is currently composing this very article within Richard Stallman’s Emacs. McCarthy’s Advice Taker left an understated but permanent influence. Warren Teitelman developed the ADVISE system in his thesis, giving it a title evocative of a pulp science fiction novel: "PILOT: A Step Toward Man-Computer Symbiosis." Teitelman later stated that he was profoundly inspired by McCarthy's initial concept of advising systems. In his thesis, he detailed a technique for halting a function call and inserting code either before or after its execution. He grew tired of manually editing other people's LISP code and sought a way to modify that code without altering it directly. He repeatedly built and refined his system, culminating in the ADVISE system within Interlisp, which served as a structured editor and runtime environment for manipulating LISP code and data. Interlisp moved away from standard text files; instead, users manipulated the data tree at a high level of abstraction using explicit commands such as “move X list to Y position.”
One can observe the author's own efforts to blend structured editing with plaintext editing in Combobulate. Eventually, AI hackers at MIT incorporated Teitelman’s ADVISE system from his PILOT thesis into their own versions of Emacs. Subsequently, it was formalized in Chalupsky’s 1992 re-implementation in Emacs Lisp, where it remains accessible today as one of Emacs’s advice systems via `(M-x find-library advice.el)`. The journey, beginning as a theoretical concept for representing automata using symbolism in the 1950
Why it matters
For working developers, this release addresses the tooling gap by aiming to create a superior coding harness beyond current AI agent interfaces. The focus on a robust terminal experience suggests an emphasis on power and efficiency for serious coding workflows. This could signal a shift in how developers interact with AI assistance.




