The Software Factory Radar

The only entry with a company behind it and the only one shipping daily. Assembly lines are DOT files extended with a model_stylesheet block - CSS-like selectors routing machine classes to different models and reasoning-effort levels, which is the most elegant idea on the radar. Machines run in Daytona cloud sandboxes with SSH and live preview into the running box, every stage takes a git checkpoint, and a REST API streams durable events. One dependency-free binary, five install paths.

Its lineage is the correction this edition makes. Fabro did not take inspiration from Attractor at arm's length; it is the successor to one of the ports (blip 11), written by the same person. Trial not Adopt purely on age - ten weeks of public life is not a track record, and the sandbox and durability claims are documented rather than independently verified here.

The most disciplined port in the set, and almost nobody has noticed it. Single compiled binary, checkpoint after every stage with resume, fan-out with fan-in consolidation, a unified client across three providers kept current with each model release, and roughly 17 reference assembly lines rather than one toy example.

The reason to assess it is a 208-test conformance suite validating behaviour against the spec. If you have ever wondered how you would write an eval for “did I implement this prose correctly”, this is the best worked answer available. No sandboxing, though - shell machines run against the host, which is blip 34.

Not on strongDM's own list, and the second-most-active port in existence. Its README states the pattern better in one line than anyone else has: it shifts the unit of durability from ephemeral agent logs to a version-controlled process graph. It is also the only port that borrows the rival family's verification idea, keeping holdout scenarios in a physically separate repository so the doer cannot reach them.

Worth watching for a reason beyond its code: ten weeks in and still accelerating is something nothing else in this quadrant managed. If it holds that pace it moves to Trial next edition. Note too that strongDM's list links a DeepCreative/dark-factory that 404s - probably this project under an old owner.

The most production-shaped port after Fabro, and the one Fabro cites as reference. Four verbs: ingest generates a DOT line from English, validate checks it semantically, run walks it, resume recovers from logs, from a purpose-built run database, or from the run's own git branch. Each machine executes in its own git worktree and commits on completion - blip 29, and the cheapest real isolation story anyone built.

On Hold because it stopped in April, but read the reason before you conclude anything about the pattern: its author did not lose interest, he built Dippin instead. This is the rare radar movement with a documented cause.

The highest-starred port that strongDM's own list omits, and Fabro's direct ancestor. A Bun monorepo of exactly three packages - unified-llm, coding-agent, attractor - which is the three-layer convergence in its purest form, arrived at independently by the person who then went and built the field leader. Archived with a redirect: “The ideas here have evolved into Fabro, a production-ready software factory built in Rust.”

Hold, obviously. Blipped because reading it beside Fabro shows a weekend port becoming a product, step by step, and because its omission from the canonical list is a lesson in itself.

Written by Justin McCarthy, strongDM's co-founder and the spec's author - which makes it the tiebreaker when the other ports disagree about what a paragraph meant. Pure C11 with no dependencies beyond libcurl and pthreads, compiling to a static library plus one binary, implementing the full spec surface with each layer mirrored by its own spec file. It even ships self-referential lines that run Attractor against its own spec to check compliance.

One commit, never touched again. Hold as software; canonical as documentation.

Four commits total, and better spec discipline than projects a hundred times its size. It ships a CONFORMANCE.md pinned to a specific upstream commit that enumerates exactly one intentional deviation and seven unimplemented upstream capabilities by name - blip 31, and the standout example of it. Five-phase execution, 13 lint rules before anything runs, and the same CSS-like per-machine model stylesheet Fabro has, arrived at independently.

Notable for honesty and for one genuinely novel mechanic. The README has a section titled “How this differs from the spec” naming its exact gaps - the manager loop is coded but not wired in, so it no-ops; resume works internally but has no CLI flag. Its fan-in machines can rank parallel candidate branches with a model prompt, falling back to a heuristic, which is the only place anyone applied model judgment to joining rather than routing. Project-doc discovery walks from git root to cwd within a 32KB budget: selective inclusion made explicit.

One blip for the long tail: the Python, Go, Kotlin, Scala, Tcl, PHP, C#, TypeScript and Elixir ports that appeared in a burst, worked, and stopped. Some are good - attractor-phoenix has 100+ commits and a real pull-request workflow on the BEAM, the most natural runtime fit anyone found, and zero stars to show for it. samueljklee is the only port that took the permission model seriously, with a shell deny-list, path confinement and injection prevention.

Together they are why the Adopt ring is nearly empty: cheap to start, cheap to demo, and abandoned the moment the author's own itch was scratched. The difference between this cohort and blip 7 is not a better idea. It is sustained investment.

Turns issue-tracker events into governed workflows: provisions an isolated workspace, injects issue context, mints scoped short-lived GitHub credentials, opens the PR, watches review and CI, tears everything down. Its own README makes the best case in this quadrant - “remote coding agents give you a shell; ElasticClaw gives you the workflow around it.” The best permission model example on the radar, and the only project treating credential scope as a first-class factory concern.

Four machine types - supervisor, worker, reviewer, merger - with full worktree isolation, dependency graphing across parallel lanes, a file-based mail system so machines can talk to each other, and a local dashboard. Cross-model review is built in rather than bolted on: model diversity used deliberately as a correctness mechanism.

Makes Gherkin acceptance criteria the contract the factory works against, with coverage tracking linking code back to business rules and a checkpoint system for safe experimentation. Its diagnosis of the problem is worth quoting to any team new to this: agents lack “the infrastructure that professional developers take for granted” - no way to force adherence to acceptance criteria, no TDD guardrails, no workflow state.

The clearest statement of this family's thesis, and honest about its lineage - it credits strongDM's Software Factory and Dan Shapiro's five-levels framing by name. Its own words: “scenarios are a holdout set. The coding agent never sees them during generation. An LLM judge scores satisfaction 0-100, not boolean pass/fail. This prevents reward hacking.” That is blip 28, put more precisely than anyone else manages.

The largest thing here by stars - an autonomous engineering-team runtime built on the AgentField platform, in public beta. Blipped mainly for scale calibration: it has more stars than every Attractor port combined except Fabro, and almost nobody discussing Attractor mentions it. The two conversations barely overlap.

The only project here that treats learning across runs as architecture. Agent traces, production logs, user feedback and team vision become Observations; a scheduler routes work items through Writer, Tester, Reviewers and a Learner; review findings return to the Writer and accumulated Expertise guides future work. Every other factory in this survey starts each run from zero.

Takes a GitHub issue and ships the merged PR, then runs scenario-driven testing through three distinct validator personas - junior, senior, maintainer. That is perspective-diverse validation rather than redundant validation: three lenses find different classes of problem, where three identical reviewers mostly agree with each other. Every run extracts patterns into a learning store.

The one project that treats a factory as something you operate rather than something you run. Its dashboard leads with execution throughput, reliability, cycle time, queue depth and worker health - manufacturing metrics, applied to software production, which is closer to the spirit of the metaphor than anything else here. Local-first: the operator API accepts loopback only.

Not a project - a warning about the search term. Two hundred and seventy-eight repositories match, and the great majority are empty scaffolds, one-day experiments, or landing pages. Several are README-only with zero bytes behind them. The name has become a genre signifier rather than a description, so treat a match as the beginning of due diligence rather than the end of it. Three candidates in this survey were eliminated by a thirty-second look at their file listing; the curated list that started this research had not done that.