Building autonomous software factories with TypeScript, Effect, and Rust — orchestrated on Kiro Crew.
I design and run multi-agent software engineering pipelines at System F Software. Human attention stays focused on intent, taste, and release, while specialized coding agents handle planning, isolated implementation, adversarial review, and rule compounding.
flowchart TD
Owner["Owner (intent, taste, release)"] --> Conductor["Conductor on Kiro Crew (requirements, plan approval)"]
Conductor --> Plan["Committed Plan"]
Plan --> Makers["Makers in Isolated Worktrees (Claude Code / OMP)"]
Makers --> PR["Pull Request"]
PR --> Reviewers["Independent Reviewers (report-only, P0-P3)"]
Reviewers --> CI["CI Gate"]
CI --> Merge["Merge"]
Merge --> Compound["Compound Step (failure becomes a rule)"]
Compound --> Conductor
The factory runs on Kiro Crew, AWS's open-source personal agent platform, orchestrating Claude Code and Oh My Pi (OMP) coding agents as makers and reviewers:
- Specialized Conductors: Dedicated conductors per product area alongside isolated ledger and security conductors.
- Editorial & Visual Loop: A Gemini copy desk, a design critic, and a page judge reviewing copy and rendered interfaces before release.
- Execution & Skill Density: Isolated workers backed by ~68 skills on the main conductor.
- Hard Guardrails & Recovery: Gate hooks that refuse unsafe actions and auto-nudge patrol loops that keep stalled sessions moving.
- Durable Memory & Deployment: Lessons that turn corrections into permanent rules, all deployed reproducibly with Nix
home-managerfrom a pinned Git ref.
Public libraries, developer tools, and conformance suites published under systemfsoftware:
| Repository | Description | Stars |
|---|---|---|
stryker-js-effect |
Effect 4 mutation testing with in-memory V8 instrumentation, Vitest runner sandboxes, and NDJSON event streams. | |
api-extractor-effect |
Effect-native API Extractor producing byte-identical *.api.md reports against @microsoft/api-extractor. |
|
are-the-types-wrong-effect |
Effect-native package export and TypeScript module resolution analyzer. | |
effect-e2e |
Effect-native end-to-end testing across core, web engine, CLI, and engine conformance suites. | |
effect-daemon |
Daemon specification and conformance suite supporting process, socket, and cluster media. | |
effect-microvm |
Microsandbox binding and daemon microVM medium for isolated execution. | |
effect-readiness |
Readiness cells and their deterministic conformance surface. | |
gritlint |
Conventions gate implemented as a Rust CLI on the GritQL engine, packaged as a Nix flake. | |
comment-checker |
Offline Claude Code hook flagging unnecessary comments and AI slop via Tree-sitter across 37 languages. | |
omp-claude-compat |
Claude Code settings hooks and CLAUDE.md @-references running natively in Oh My Pi agents. |
|
constitution |
Engineering constitution (functional core / imperative shell, types-first, mutation-tested) vendored across repos and enforced by AI coding agents. | |
hex-schema |
Hex-encoded byte schemas built for Effect Schema. | |
rx-effect |
RxJS Observable to typed Effect Stream bridge with backpressure and interruption semantics. |
|
effect-memfs |
In-memory filesystem implementation for Effect (fork of nounder/effect-memfs). |
|
discern |
Uncertainty-aware semantic pattern matching and control flow for Effect (fork of doeixd/discern). |
|
effect-endgame-starter-kit |
Opinionated project template pre-wired with Effect, strict typing, mutation testing, and agent guardrails. |
- Plan before code: An outer-loop conductor owns requirements and approves a committed plan in the repository before a single line of implementation is written.
- Inspect the state, never the transcript: Automated checks read only the produced repository state and artifacts, never how an agent arrived there.
- Reviewers report, CI decides: Independent reviewers audit the exact PR head in report-only mode (
P0–P3), while deterministic CI remains the sole merge gate. - Taste belongs in review: Aesthetic, architectural, and API design calls go to human or critic review, never to brittle lint rules.
- State lives in Git and the issue tracker: Ephemeral chat context is disposable; durable memory lives in committed plans, rules, and issues.
- Done means verified in production: A task is complete only after it is deployed and checked against the real system.
Get the long-horizon prompt to build your own software factory at systemfsoftware.com.
Ranked Top 50 Grandmaster in StarCraft II during the Heart of the Swarm era.
o





