deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

Mastering Emacs author releases DreamCoder, a browser-based AI coding harness with a tiling window manager

After fifteen months of development, the author behind Mastering Emacs has released DreamCoder, an AI coding harness with a tiling window manager that runs in the browser on all three platforms.

Mastering Emacs author releases DreamCoder, a browser-based AI coding harness with a tiling window manager

A harness built around tiling windows, not a chat pane

The author of the Mastering Emacs blog has announced DreamCoder, an AI coding harness released after fifteen months of work and a beta test. According to the announcement, which reached the front page of Hacker News, DreamCoder is what he calls a "frontier coding harness" built around a tiling window manager, and it runs in the browser on all three platforms. The Hacker News listing describes it as both a coding harness and an IDE.

Why build another coding agent

The essay opens with the author's first encounter with Claude Code, which he installed a few days after its first release in May 2025. He set it loose on a complex piece of enterprise middleware he had built several times over his career, and was struck by the pace: the agent jumped between disparate languages, frameworks and theory from one turn to the next, even though it demanded constant human steering and correction.

What bothered him was the packaging. Claude Code was the first terminal application he had used in a long time — an odd choice for a coding agent, in his view, and one the rest of the industry promptly copied. GUI front-ends exist, he writes, but they are second-class citizens next to the terminal, and most amount to a boring sidebar with a chat pane. Even casual users bristle, he notes, when they try to copy multi-line text in a terminal.

So in May 2025 he decided to build a better harness, reasoning that with AI assistance it should take two or three months. Fifteen months later — after what he describes as endless UI and backend iterations and a successful beta — DreamCoder shipped. That timeline doubles as a wry data point about AI-assisted development itself.

A potted history of AI and Emacs

Before getting into the product, the essay traces Emacs's and LISP's place in the history of AI. In the 1950s at MIT, John McCarthy set out to build the Advice Taker, a system meant to encode commonsense reasoning in mathematical logic: if ordinary facts about life could be expressed with symbolic precision, the rudiments of AI might follow. Formalising such a system required a language where code and data could intertwine recursively, and ordinary programming languages would not do.

That language was LISP, designed with two facets: M-expressions, an Algol-like functional notation using square brackets, and S-expressions, a parenthesised format for symbolic data that was never intended for writing code. The specification's eval feature implied the language was itself a universal Turing machine, but McCarthy lacked the time, staff and money to build the recursive evaluator. Steve Russell, who worked with him, proved the point by simply doing it, writing the first LISP interpreter with a working eval. M-expressions were dropped as a matter of pragmatism and survive today mainly in Wolfram's Mathematica. McCarthy later conceded that "LISP owes its survival specifically to the fact that its programs are lists, which everyone, including me, has regarded as a disadvantage."

The Advice Taker itself never fully materialised. Its chains of small rules and facts — the essay shows snippets like (at me home) and (walkable home car) — anticipate Prolog, but the axioms never stopped multiplying: yes, you can walk to your car, but what if the front door is locked, or the car is parked on the street? The author illustrates the problem with a teenage internship at a banking-software firm, where teams were told to flowchart the assembly of a bike as explicitly as possible; he lost because his chart forgot to open the box the bike came in.

McCarthy moved on to found an AI lab at Stanford, but the idea lived on. In the late 1970s, the essay recounts, Richard Stallman and Guy Steele created EMACS — Editing MACroS — at MIT's AI Lab, as a set of building blocks on top of TECO, a famously inscrutable editor that did not even refresh the screen by default.

Why it matters

The current generation of coding agents has converged on a remarkably narrow interface: a terminal session, or a chat sidebar bolted onto an editor. DreamCoder is a deliberate counter-proposal from someone steeped in the Emacs tradition, where the editing environment is a programmable, composable workspace rather than a conversation window. The essay's history lesson makes the argument explicit: Emacs and LISP are not incidental trivia but a founding strand of AI tooling, and today's harnesses are the latest chapter of that lineage. The fifteen-month build, set against an original two-to-three-month estimate made by a developer using the very AI tools he was building for, is also a useful reality check on how much human steering AI-assisted development still needs.

  • #ai-coding
  • #emacs
  • #developer-tools
  • #ide
  • #lisp

Related posts