Posted February 08, 2026

The novel is called Metacompiler. It's available as a free PDF download and on Amazon.

I collaborated with Claude Code's new AI agent teams feature to finish and publish my novel in under 24 hours.  This is how I made that happen.

The Idea That Wouldn't Die

In 1984, Ken Thompson delivered his Turing Award lecture, "Reflections on Trusting Trust." In six pages, he demonstrated that a compiler could be modified to insert a backdoor into any program it compiled — and then modified to insert that modification into the next version of the compiler itself. Even if the original malicious source code was deleted, the backdoor would propagate invisibly through every future compilation, forever.

Thompson's paper is taught in computer science programs around the world. It is elegant, concise, and deeply unsettling. The theoretical foundation is real, and variants have surfaced in the real world — the SolarWinds compromise in 2020, the XZ Utils backdoor in 2024 — each one a reminder that software supply chain attacks are not academic exercises.

I have spent thirty years working with embedded software: the firmware in your car's throttle controller, your pacemaker, your insulin pump. In 2013, I served as the lead software expert witness in a Toyota unintended acceleration trial — the only such case to reach a jury. My team spent eighteen months examining millions of lines of engine control code. The jury found Toyota liable.

That experience planted a question I could not shake: What if vulnerabilities didn't exist in individual product designs, but instead in a compiler that built every product's code?  If you asked me to defend any technical claim in the novel under cross-examination, I could do it. I've done it before.

For years I accumulated notes. Character sketches and draft chapters. Research files on compiler design, signals intelligence, medical device vulnerabilities. I fell in love with the story arc as well as the main character. I possessed and provided the technical expertise to invent such a plot. What I did not have was the ability to sit down and write a 100,000-word novel. A novel is not a technical specification. The skills don't transfer.  And finishing this was never a priority I could find enough time to complete my own.

AI Agent Team Architecture

On February 7, 2026, I opened Claude Code (Anthropic's command-line AI tool) and built a writing system based on the prior day announcement of experimental agent teams.

The architecture was the key:

  • Claude Code served as the orchestration layer — a command-line interface that could spawn and coordinate cooperating AI agents to do their work in parallel.
  • Three writers each assigned a different average chapter length but given the same story bible based on my prior thoughts and ideas were paired with
  • Three editors each assigned to perform the chapter-by-chapter reviews of one author
  • Four editorial reviewers evaluated individual act-length drafts as well as the overall work from the perspective of different likely readers:
    • Jake — thriller pacing and tension
    • Sarah — character depth and emotional authenticity
    • Marcus — thematic consistency and theological subtext
    • Dr. Chen — technical accuracy of the computer science
  • A quality gate: every chapter had to receive an A-tier grade from all four reviewers or get rewritten.
  • A publisher who worked to review best practices for publishing hard sci-fi on Kindle at present and weigh the perspectives of the reviewers' feedback on the competing drafts against that criteria and to bring me the difficult final decisions
  • A marketer who generated social media post ideas, such as this one, in concert with the evolution of the book itself.

What the Agents Did Well

Voice consistency. Across 43 chapters and multiple point-of-view characters, the prose maintained a unified style. This is harder than it sounds. A human author writing over multiple sessions will drift. The agents, given a style reference and character voice profiles, stayed on target.

Continuity tracking. The novel has specific technical details that must remain consistent: the number of nodes in a distributed computing mesh, character ages, geographic distances, the hex commands in a CAN bus exploit. The review panel caught real problems — a character's age didn't match the timeline in one chapter, a location was wrong for the narrative geography in another.

Speed. A 2,500-word chapter drafted, reviewed by an editor while his author moved on to the next, a publisher and marketer thinking about what could work well with readers, and sample reader feedback available nearly instantly for consideration. On average, the final chapters were drafted, edited, reviewed, reconsidered, and finalized in an average of fifteen minutes.  Claude Code made it easy for me to change a detail and have it correctly tweaked and the implications propagated across the entire story arc in a minute. 

What I Had to Do

Every creative decision was mine.

Inspired by a book I'd once read about crafting a good novel I'd thought deeply about the story framework. I designed each character's background, motivations, and evolutions. I chose their relationships as well as who lived and who died (and how).  I documented the technical details underlying the vulnerability and its potential uses in attack and counterattack: the three-command backdoor (INFO, PEEK, POKE), the distributed mesh architecture, the medical device attack vectors, the compiler genealogy from Bell Labs to the present day. I wrote chapter-by-chapter outlines and assembed the details that became the story bible passed to the agents to do their tasks.

And I caught things the agents missed and misunderstood. A geography error that put a character on the wrong  highway. An assumption about the narrator's relationship to Kali that contradicted the story's premise. A  fire suppression system that had been banned in new installations since the Montreal Protocol. These are the kinds of details that come from domain expertise, not language modeling.

The AI agents and I collaborated on this project. They generated and refined prose within my framework. The best analogy is architecture: I architected this book and selected amongst design options. The agents performed heavy lifting and carpentry using their tools. 

The Meta-Narrative

Here is the part I find genuinely interesting. (Spoiler alert.)

METACOMPILER is a novel about how a backdoor hidden in compiled code propagates through every device on earth. The protagonist, a deaf-blind computer scientist named Kali Devi, discovers the backdoor, builds a distributed supercomputer to counter it, and ultimately creates the first artificial general intelligence — as an act of love, designed to ask permission before it acts.

That AGI narrates the entire novel. The reader doesn't know this until the epilogue.

The meta-narrative is deliberate: the same class of technology the novel describes — machine intelligence operating within human-defined constraints — is the technology that helped me write it. A novel about trusting code, written with the assistance of code. The question Thompson posed in 1984 (“To what extent should one trust a statement that a program is free of Trojan horses?”) applies to this novel itself.

I don't think this makes the book less authentic. I think it makes the book's argument more honest. The AI that helped write METACOMPILER operated under my direction, within my framework, subject to my editorial judgment — exactly the relationship between Kali and the AGI she creates in the story.

Read It

METACOMPILER is available now:

The copyright page includes a full disclosure of the AI collaboration. I believe transparency about how a book was made is a prerequisite for the reader's trust — which is, not coincidentally, what the novel is about.

The backdoor is fiction. The vulnerability is real. Thompson's six-page paper should terrify everyone who depends on compiled software, which is everyone. I wrote a 101,000-word novel to make sure it does.

Michael Barr is the CTO of Barr Group and one of the world's foremost experts on embedded software. He served as the lead software expert witness in a Toyota unintended acceleration trial and is the author of three technical books. METACOMPILER is his first novel.