What if the point of an AI coding tool wasn’t to help you type faster, but to make the keyboard unnecessary? AgencyX takes an engagement from a GitHub issue all the way to a merged pull request on your own Apple hardware, and leaves a durable, replayable trail of evidence at every step.

The conversation covers:

• Sovereign compute, auditable evidence. It runs on your machine, in your language (Swift), and every model call and tool action lands in a replayable event ledger you can audit.
• The lights-off thesis. The human directs and reviews at gates; the agency does the work in the dark. Its real success metric is operator silence.
• A layered Swift architecture tuned for local hardware, with a self-improving flywheel that uses fine-tuned local models for the routine work.
• Orchestration under the hood: configurable workflows, context compression to cut cost, and a decomposition cycle (refined from its predecessor, loswfx) that sizes work correctly before a line is built.
• Strict safety gates: how a system that builds and ships its own milestones still stops for the human at exactly the moments that matter.

This episode is a machine-generated deep dive, built from AgencyX’s own design and research documents. It pairs with the companion short, Hiding the Messy Mind of AI.

The agency does its real thinking, the false starts, the research, the discarded plans, in a private shadow brain kept out-of-tree, so your repository only ever sees the clean, durable result. The mess stays hidden; the evidence stays honest.

A visual companion to Why AgencyX Banned the Keyboard.

The conversation walks the pipeline:

• Text in, structure out. The writer works in Fountain or Final Draft; khaos-parser ingests the raw text and generates the KSPD (the Khaos Story Project Data), the single source of truth for the story model.
• Separation of concerns. Text is for the writer; structure is for the machine. The parser is the only thing that writes into the story model.
• Identity that persists. Every scene and beat gets a ULID, so when scenes move, their data moves with them.
• Read, don’t rewrite. story-graph and story-agent read the KSPD to reason about the narrative; the writer feedback loop (khaos-wfl) watches it and appends signals to a side file, never overwriting the source.
• Local-first. Privacy-centric and offline-capable by design.

Recorded January 2026, generated from the Khaos Machine design notes.

Topics:

• Parsing speed. Turning a full screenplay into structured data quickly enough to run as the writer types.
• Production concerns. What it takes to move a parser from a prototype to something dependable.
• ULIDs as identity. Every scene, beat, and element gets a sortable, stable identifier, so when the script is reordered or revised the machine still knows what moved where, and the analysis stays attached to the right thing.

Recorded January 2026, generated from the Khaos Machine design notes.

A short video overview on how a modern script gets made: the loop of drafts, coverage, and revision that carries a story from concept to production.

The overview walks through:

• The journey, not the draft. A screenplay is a thing under development, not a finished object. The interesting question is how it moves.
• Feedback and revision as the core loop. Coverage and notes are the engine. The work is turning conflicting reactions into the next draft.
• Development as diagnosis. The overview leans on a medical analogy: the concept is the patient’s history (qualitative), the coverage is the blood work and X-rays (quantitative), and the producer is the lead physician who has to reconcile contradictory specialist opinions (marketing, director, talent) into one clear treatment plan. Five conflicting prescriptions and the patient gets sicker (development hell); one synthesized cure and the story recovers (production).
• Workflow and tooling. Where the technology fits into that loop, and how a feedback-driven ecosystem keeps the writer’s words intact while giving everyone else the structure they need.

Recorded December 2025, generated from notes on modern, feedback-driven screenwriting.