AI software factories: a survey of backlog-to-merged-code systems

Which AI software factories exist: systems where a backlog or spec goes in and verified, merged code comes out with minimal human involvement? AR Labs carried out this survey in September 2026 for our own AI software factory project. It compares more than 60 open-source, commercial and internal-company systems using GitHub repository statistics, README checks and web searches, and rates the leading twelve against each stage of the pipeline.

A software factory here means an end-to-end system: a backlog or spec goes in, and verified, merged software comes out with minimal human involvement. The stages are intake, spec, decomposition, parallel implementation, verification, review, merge queue, deploy, and fleet monitoring with cost control. This document surveys what exists, weighted toward what a solo operator with a backlog in Beads (bd, Steve Yegge's issue tracker for coding agents) and Claude Code/Codex could run with 5-10+ agents.

Method and limits

• Web searches: 9 queries (WebSearch). The Bright Data batch search returned errors, and the researcher query budget capped breadth, so the planned target of 20 queries was not reached. Coverage gaps are listed at the end.
• GitHub: 15 gh search repos queries sorted by stars (the first attempt hit the API rate limit and was re-run after reset), plus gh api stats for 65 repos and contributor counts for 18.
• The single most productive source was the curated list varun1505/awesome-software-factories, which catalogues about 390 entries across factories, ticket-to-PR pipelines, loops, session managers, isolation, spec layers, review gates, and commercial products. It was extracted in full, and stars were cross-checked on the leading candidates.
• README verification: 23 READMEs were downloaded and grepped for Beads, merge queue, worktree, holdout, judge, deploy, cost, and agent names. Eight were read in more depth.

Stars and dates are as of 2026-09-28. "Created" is the repo creation date, and "Push" is the date of the last push.

(a) Longlist

Established reference systems

Name	Type	Stars	Created / Push	Contributors	Notes
Gas Town	OSS (Go)	18,206	2025-12 / 2026-09-18	n/a	Yegge. Mayor dispatches to parallel agents. Beads is the data plane; Refinery merge queue.
Gas City	OSS (Go)	1,309	2026-02 / 2026-09-28	203	Gas Town recast as an SDK for building dark factories; v1.0. gascity.com
OpenRig	OSS (TS)	1,530	2026-04 / 2026-09-28	14	YAML-defined team of Claude Code and Codex agents over tmux.
StrongDM Attractor	Spec (NL spec)	1,320	2026-02 / 2026-03	n/a	NL spec of a non-interactive coding agent for a factory. Scenarios as holdouts, Digital Twin Universe.
Factory.ai	Commercial	n/a	n/a	n/a	"Droids" across the SDLC. $200M at a $5B valuation in 2026. Pro $20/mo, Plus $100, Max $200.
Sandcastle	OSS (TS lib)	8,170	2026-03 / 2026-06-29	12	Orchestrates agents in Docker/Podman/Vercel sandboxes plus worktrees. Push activity has slowed.
OpenAI Symphony	OSS spec + Elixir	27,457	2026-02 / 2026-09-15	9	Issue-tracker items become isolated autonomous Codex runs. Ports: Contrabass, Sortie, oh-my-symphony.
Stripe Minions	Internal	n/a	n/a	n/a	See section (c).

Newly discovered: open-source factories (end-to-end or close to it)

Name	Stars	Created / Push	Contrib.	What it is
Fabro	1,663	2026-03 / 2026-09-28	26	Rust single binary, "open source dark software factory". Workflows are Graphviz DOT graphs with loops, parallelism, and human gates. CSS-like model routing across vendors. Daytona cloud sandboxes. Verification nodes (tests, linters, LLM-as-judge) with fix loops. Git checkpointing and a REST/SSE API.
Optio	1,054	2026-03 / 2026-09-26	14	Self-hosted platform. Tickets from GitHub, GitLab, Linear, Jira, or Notion become merged PRs. It watches CI, runs a blocking review agent (which can use a different vendor), and uses a Kubernetes-style reconciliation control plane. Runs on K8s pods or local machines. Supports Claude Code, Codex, Copilot, Gemini, Cursor, and OpenCode. Tracks cost.
Fusion	1,249	2026-04 / 2026-09-27	46	Multi-node orchestrator with a kanban board, plan-review-execute gates, per-task worktrees, and a large dashboard.
loki-mode	1,073	2025-12 / 2026-09-28	7	PRD or issue in, PR or deployed product out. Claims 41 agents in 8 swarms, 9 quality gates, blind review, and a cost dashboard. The claims are large and unverified.
SWE-AF	1,015	2026-02 / 2026-09-24	14	"Autonomous SWE fleet": plan, code, test, and ship a PR through one API call. Worktrees, Docker, budgets. Supports Claude, Codex, and OpenCode.
zeroshot	1,915	2025-12 / 2026-09-28	16	Planner and implementer loop with independent validators, in worktrees or Docker. Claude and Codex.
AI Software Factory (on Archon, 23.6k stars)	337	2026-09 / 2026-09-28	1	GitHub issue goes through mission triage, plan, build, independent judge, holdout verification, merge queue, and deterministic deploy. Scheduled regression runs re-test merged work. This is the most complete stage coverage found, but the repo is new.
software-factory-rs	1	2026-08 / 2026-09-03	1	Beads-native. A Planner turns plan text into a Beads epic of task and verify beads. Stateless pull-based workers claim ready beads under leases. A non-LLM verifier runs checks in a clean worktree and then merges on the rebased result. Rootless Docker with default-deny egress. Claude Code, OpenCode, or Codex. Treat it as a reference architecture rather than a dependency.
Beadhive	22	2026-07 / 2026-09-28	1	bh: multi-repo Beads orchestration ("Agentic Git Flow" implemented on Beads), MCP-native.
eve Software Factory Template ("Foreman")	1,138	2026-08 / 2026-08-20	n/a	Vercel Labs template in which an agent owns each SDLC stage, with Linear intake and deploy.
Super Simple Software Factory	903	2026-08 / 2026-08-04	n/a	Deterministic Python owns the workflow graph and agents are bounded steps. Sibling project Inkwell runs it on throwaway VMs.
Atomic	834	2025-10 / 2026-09-27	37	"Verifiable coding agent runtime": the process is a natural-language stage graph with checks.
dzhng/skills	958	2026-07 / 2026-09-27	n/a	Skills for factories: explore, spec, implement, review, and run autonomously.
DoorDash agentic-orchestrator	108	2026-05 / 2026-09-25	8	DoorDash's deterministic harness: feature description goes through gated phases to a reviewed PR. Claude, Codex, OpenCode.
Taskplane	217	2026-03 / 2026-09-19	n/a	Multi-agent orchestration, self-described as "light factory". Worktrees and a dashboard.
dark-factory-experiment	162	2026-04 / 2026-08	n/a	A repo that ships its own code: triage, implement, review, and auto-merge.
OctopusGarden	64	2026-02 / 2026-03	n/a	StrongDM-style: specs in, code out through a convergence loop, holdout scenarios, and a judge. Stale since March.
DUBSOpenHub/dark-factory	27	2026-02 / 2026-09	n/a	"Sealed-envelope" testing: agents cannot see the tests.
fspec	95	2025-10 / 2026-09-27	n/a	Spec-driven multi-agent harness for dark factories (Rust).
Sgai	133	2026-01 / 2026-09-21	n/a	Sandgarden's local factory: a single GOAL.md drives a coordinated agent set.
CodeAF	131	2026-08 / 2026-09-28	n/a	Single binary that splits tasks into parallel branches and auto-checks and merges them.
Mastra Factory	47	2026-07 / 2026-09-25	n/a	Factory server connecting GitHub, Linear, and Slack with configurable workflows.
Miniforge	44	2026-01 / 2026-09-28	n/a	Policy-as-code governance over autonomous development (Clojure).
Machinist	468	2026-07 / 2026-09-28	n/a	Worker-fleet infrastructure that pulls tasks into isolated workspaces.
Foreman (VisionForge)	440	2026-06 / 2026-06	n/a	TUI supervising headless Claude Code agents through a gated pipeline.
Genie	343	2025-08 / 2026-09-28	n/a	Interviews, plans, and dispatches parallel agents to worktrees, then reviews.
addyosmani/factory	211	2026-08 / 2026-08	n/a	Issues as a work queue. Scheduled agents triage, implement, test, and review, then open draft PRs.
Helmor	1,307	2026-04 / 2026-08	n/a	Local workbench: plan, run, review, test, merge, and ship in worktrees.
Fluent	107	2026-07 / 2026-09	n/a	Vision docs, bug reports, and production logs in, working software out, with self-improvement.
gstack	134k	2026-03 / 2026-09-28	n/a	Garry Tan's Claude Code setup branded as a "software factory". It is a skill pack, not an orchestrator.
Paperclip	92k	2026-03 / 2026-09-28	219	Orchestration for "zero-human companies": org charts, budgets, cost tracking. Broader than code.
Multica	51.6k	2026-01 / 2026-09-28	326	Self-hostable workspace where agents take issues like teammates.
Loki/Symphony ports: Sortie, Contrabass, oh-my-symphony	193 / 222 / 30	2026	n/a	Symphony pattern in Go and Python with multi-agent CLI support.

Ticket-to-PR pipelines (partial factories: intake through PR, weak on merge and verify)

Open SWE (10.8k stars; LangChain's open-source version of the Stripe/Ramp/Coinbase pattern), OpenHands (89k), background-agents (3.3k), GitHub Agentic Workflows gh-aw (5.2k), Cyrus (832), centaur (1.3k, Paradigm), Mission Control (6.3k), no_human (324), groundcrew (66, Clipboard Health), Yak (33, Geocodio), ccpm, and AWS remote-swe-agents.

Loops (single-lane autonomous runners)

ralph-orchestrator (3.2k), continuous-claude (1.4k), CodeMachine-CLI (2.5k), nightshift, ghuntley/loom.

Commercial and hosted factories

Product	What it does	Pricing / solo usability
Factory.ai	"Factory 2.0: from coding agents to software factories". Droids across the SDLC.	$20 / $100 / $200 per month. Usable solo.
Warp Factories / Oz	Prebuilt loop: triage, spec, implement, review, verify, ship, monitor. BYO model.	Early Access for teams. Pricing not public.
Blitzy	Enterprise factory running thousands of agents, aimed at modernisation.	$80/mo base + $40/seat + compute units. Enterprise-oriented.
8090 Software Factory	Governed full-SDLC workspace for people and agents.	Enterprise.
Autoheal	"Self-improving software factory" for post-coding work: incidents, vulnerabilities, cost. $7.9M seed on 2026-09-28.	Enterprise, runs in the customer's cloud.
Devin	Autonomous engineer with parallel worker instances.	Free tier; Pro $20; Teams $80 + $40/seat.
LightSprint, Charlie Labs, Tembo, Augment Cosmos, Amika, Hoplite	Hosted parallel-agent factories or daemons.	Varies. Not verified.
Linear coding sessions, GitHub Copilot coding agent, Cursor Cloud Agents, Codex cloud, Claude Managed Agents, Jules, Kiro, GitLab Duo Agent Platform, Atlassian Rovo Dev	Cloud agents: issue to PR. Implementation lanes, not factories.	Per-seat or usage.

Total distinct factory or near-factory systems in this document: more than 60.

(b) Stage-coverage matrix: top 12

Key: Y = built in, P = partial or needs configuration, N = absent. Ratings come from README evidence. Gas City and OpenRig ratings also draw on prior knowledge and carry moderate confidence.

System	Intake	Spec	Decompose	Implement (parallel)	Verify	Review	Merge queue	Deploy	Monitor / cost	Agents	Isolation	Beads?
Gas City	Y (Beads)	P	Y (formulas/molecules)	Y	P	P	Y (Refinery)	N	Y	Claude, Codex, Gemini	worktrees/tmux, Docker opt.	Native
Fabro	P (API queue)	Y (NL specs)	P (hand-authored graph)	Y	Y (tests, lint, LLM judge, fix loops)	Y (cross-model critique)	P (git branches)	N	Y (events, model-routing cost)	Any vendor	Daytona VMs	Adapter needed
AI Software Factory (Archon)	Y (issues + mission triage)	Y	P	Y	Y (holdout + runtime)	Y (independent judge)	Y	Y	P (scheduled regression)	Configurable	Archon worktrees	Adapter needed
software-factory-rs	Y (plan text)	Y (verify-first)	Y (Beads epic)	Y (pull, leases)	Y (non-LLM verifier)	P	Y (rebased checks)	N	P (budget)	Claude, OpenCode, Codex	rootless Docker, deny egress	Native
Optio	Y (GH/GitLab/Linear/Jira/Notion)	P	P	Y	Y (watches CI)	Y (blocking review agent)	Y (to merged PR)	N	Y (cost)	6+ CLIs	K8s pods / local	Adapter needed
Symphony (+ ports)	Y (Linear)	P	N	Y	P	P	P	N	P	Codex (ports: many)	per-issue workspace	Adapter (spec invites it)
OpenRig	P	P	P	Y	P	Y (peer review)	N	N	P	Claude + Codex	tmux	Unknown
Sandcastle	P (issues)	N	N	Y	P	P	P (template auto-merge)	N	N	Claude, Codex, OpenCode	Docker/Podman/Vercel + worktrees	Mentioned in README
zeroshot	P	Y (planner)	P	Y	Y (validators)	Y	P	N	P	Claude, Codex	worktree/Docker	No
SWE-AF	P (API)	Y	Y	Y	Y	P	P (ships PR)	P	Y (budget)	Claude, Codex, OpenCode	worktree/Docker	No
loki-mode	Y (PRD/issue)	Y	Y	Y	Y (9 gates)	Y (blind)	P	Y (claimed)	Y	Claude, Codex, Gemini, OpenCode	worktree/Docker	No
Fusion	Y (kanban)	P	P	Y (multi-node)	P	Y (review gate)	P	P	Y (dashboard)	Claude, Codex, Gemini	worktrees	No

No system covers every stage well. Deploy and a true merge queue (serialised rebase, re-test, then land) are the rarest stages. Only Gas City (Refinery), Archon's SDLC pack, and software-factory-rs implement a merge queue explicitly.

(c) Internal-company reference architectures

Three companies (Stripe, Ramp, Coinbase) converged on the same shape independently, and LangChain's Open SWE packages it. The shared pattern: invocation from chat or a ticket, a pre-warmed isolated sandbox, org context injected, deterministic gates wrapped around agent loops, a PR, and a human merge.

Company	System	Architecture points worth copying	Scale claim
Stripe	Minions	Slack or ticket trigger. Pre-warmed EC2 devboxes (about 10 s to provision), isolated from production and the internet. The agent runs against a subset of the 3M-test suite before opening a PR. Deterministic steps are interleaved with agent steps.	about 1,300 PRs/week (podcast, Ona white paper)
Ramp	Inspect	Background agent on sandbox infrastructure, with a full dev environment per session.	over 50% of merged PRs (Ona)
Coinbase	Cloudbot	Same convergent pattern.	n/a
Spotify	Honk	Claude Code as the harness. Slack in, build status out to Slack, merge from phone. Aimed at fleet-wide migrations, with verification loops before a PR.	1,500+ merged PRs; migrations 60-90% faster (Spotify Eng)
DoorDash	Flux, plus the open-source agentic-orchestrator	Platform primitives: agent sandboxes, MCP gateway, playbooks, invocation surfaces. Gated phases.	130k tasks/month; 25k automated reviews/week (DoorDash)
Block	Builderbot (also Goose, buzz)	Agents as channel members with audit trails.	about 1,500 merged PRs/week, about 15% of production changes (secondary source)
Shopify	Roast + auto-review	Ruby DSL for structured workflows. Auto-review and merge of low-risk PRs.	60-70% of low-risk PRs merged with no developer time (secondary source)
Airbnb	n/a	Agentic sessions at org scale.	64% of PRs (secondary source)
StrongDM	Software Factory / Attractor	No human reads the code. Specs plus scenarios stored outside the repo as holdouts. A "satisfaction" metric replaces pass/fail. A Digital Twin Universe of behavioural clones of dependencies enables high-volume validation. (Simon Willison)	Level-5 "dark factory"
HumanLayer (Dex Horthy)	Cautionary tale	A fully lights-out factory run from Jul-Nov 2025 degraded the codebase until one bug took weeks to fix manually (BigGo).	Argues for keeping humans on specs and architecture

A research corpus of 20+ company case studies exists at Chipagosfinest/software-factory. It was not read in depth.

(d) Ranked top 6 for a solo operator with a Beads backlog

The ranking assumes a solo operator, a Beads backlog, Claude Code and Codex, 5-10+ agents, and a need for verified merges rather than PR spam.

1. Gas City (repo). It is the only mature, actively developed factory whose control plane is Beads. It has a merge queue (Refinery), 203 contributors, and v1.0. It consumes an existing Beads backlog with no adapter. Its costs are conceptual weight, coupling to Beads versions, and weak verification and deploy stages, which the operator would have to add.
2. Fabro (repo). It has the strongest verification and model-routing layer: a DOT workflow graph, LLM judges, fix loops, cross-vendor critique, and a single Rust binary. It follows the StrongDM lineage. Consuming Beads needs a thin adapter that feeds bd ready into Fabro runs. A good pairing is Gas City or a Beads dispatcher on the outside, with Fabro workflows as the per-bead verification pipeline.
3. software-factory-rs (repo). Its architecture is the closest match to the target described above: Beads-native, pull-based workers with leases, verify beads planned before code, a non-LLM verifier, rebased merge checks, and a rootless deny-egress sandbox. It has 1 star and 1 contributor, so copy its design rather than depend on it.
4. AI Software Factory on Archon (repo). It has the widest stage coverage found: holdout verification, an independent judge, a merge queue, deploy, and scheduled regression. Archon itself is mature (23.6k stars), but this wrapper is days old and intake is GitHub Issues, so Beads would need a sync step.
5. Optio (repo). It is the most polished ticket-to-merged-PR loop: CI watch, a blocking review agent on a different vendor, six agent runtimes, cost tracking, and a local mode. It needs a Beads-to-tracker adapter, and its full mode is Kubernetes-heavy.
6. Symphony spec, with the Sortie or Contrabass ports (repo). It is a vendor-backed open spec for tracker-to-isolated-run orchestration, so writing a Beads tracker adapter against a spec is well-defined. Coverage beyond implementation is thin.

Also worth watching: Beadhive (multi-repo Beads flow), zeroshot (independent validators), DoorDash agentic-orchestrator (a production-derived gated harness), and OpenRig (a team runtime for Claude and Codex, weak on merge and verify).

Gaps not covered

• Web query count fell short of the planned target of 20, because Bright Data errored and the query budget applied. Some YC-batch factory startups are likely missing.
• The "updated" sort for GitHub searches was not re-run after the rate limit hit.
• Commercial pricing was verified only for Factory, Devin, and Blitzy.
• Internal scale figures for Block, Shopify, and Airbnb come from secondary sources.
• Gas City and OpenRig stage ratings were not re-verified from their READMEs in this pass.