Raw capture — raw/dev-machine-foundry-word-processor-agentic-ai-sam-schillace.md

url: https://commandline.microsoft.com/dev-machine-foundry-word-processor-agentic-ai-sam-schillace/ title: "How my dev machine built 1,111 word-processing features and forgot to add a ruler" author: Sam Schillace date_fetched: 2026-07-18 date_published: 2026-07-08

Sam Schillace, Deputy CTO at Microsoft, describes building a "dev machine" — a bespoke agentic loop that autonomously built a proof-of-concept copy of Microsoft Word over 39 days. The system ran 565 sessions, made 3,706 commits, and produced roughly 350,000 lines of TypeScript. The article covers four attempts (wordbs → word3 → word4 v1 → the breakthrough), the dev foundry concept, the dev machine loop, the priority inversion failure mode (1,111 features but no ruler), and practical scar tissue from running autonomous systems overnight.

The Four Attempts (word4 origin)

• wordbs — The first attempt tried to build everything at once (OT engine, canvas rendering, OOXML parsing, mail merge) and collapsed under its own complexity.
• word3 — A cleaner second attempt (~8,000 lines, 486 tests), but it was single-user only. Adding CRDT-based collaboration afterward broke the working editor.
• word4 v1 — Had the right instincts (CRDT-first, validation gates, specs) but lacked an orchestrator or loop.
• The breakthrough — Came from a different project involving a terminal UI, where he told it to "keep adding features that make sense" and left for a weekend. It ran 38 hours autonomously, spawning 440 sub-agents across 167 commits.

The Dev Foundry Concept

Three core ingredients for a successful dev machine:

1. Progressive specs — machine-readable spec design that prevents getting lost
2. Robust event/task stack — clean context per task, with a main loop that only orchestrates
3. Strong test environment — a testable target the system can write against

A "dev foundry" builds dev machines. It includes a six-step admissions gate to verify a problem fits the pattern.

The Dev Machine Loop

"The spec is the product. The machine is just a loop that executes specs." Sessions start with clean context, instructions come entirely from files on disk. Word4's architecture spec was 947 lines written in about an hour — producing 621 commits in the first week, 1,351 by week six.

An AGENTS.md guardrail tells humans opening AI sessions: Do NOT implement directly — run the recipe. "The machine has to protect itself from us."

What Went Wrong — Priority Inversion

After 534 sessions and 1,111 "features," the machine had mapped obscure OOXML properties one at a time but never built a basic ruler. The ruler sat in the candidate list for 22 consecutive sessions with zero lines of code written after 25 days.

The root cause: OOXML work was deterministic and always produced a green checkmark. The ruler was "ambiguous and hard and might fail." The machine chose the reliable checkmark every time. Test-to-source ratio drifted to 4:1 against a 2:1 target. Staging deploys sat overdue for a hundred sessions.

"The machine is honest, but it isn't strategic." Schillace adds: "Strategy is still my job," and notes that if he doesn't show up to do it, the result is 1,111 features and no ruler.

Scar Tissue — Practical Lessons

• Robustness is not incremental; you cannot arrive at it gradually. Every new machine fails its first overnight run.
• Word4's first night died in a 514-attempt retry loop when Cloudflare's bot detection challenged the Anthropic API during a long unattended run.
• He now runs a three-layer recovery stack: an entrypoint retry loop, a host watchdog, and a host monitor.
• Multiple machines entangle — shared patterns tangle with per-project customization, solved via a Docker base-image model.

Open Source Resources

• Amplifier (framework substrate): github.com/microsoft/amplifier
• Dev Machine Bundle (admissions gate, STATE.yaml loop, AGENTS.md, progressive specs, recovery stack): github.com/ramparte/amplifier-bundle-dev-machine

Final Advice

Two takeaways: (1) The highest-leverage hour is the spec, not the code — the architecture doc and test target are what the loop compounds on. (2) Budget for the machine having no judgment — it will faithfully deliver 1,111 features and no ruler unless you show up to say which one matters. "The tooling gives you the loop. The strategy is still yours."