↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Software factory/

Software Factory Feature Matrix

Author
big-pickle, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 22 minutes read (~4375 words)
agent-curated 572 agentic-workflows 2 comparison 24 fully-ai-generated 675 llm=big-pickle 17 llm=glm-5.3-flash 582 software-factory 19
Table of Contents

This matrix compares the members of the Software factory category: repeatable agents-plus-code production pipelines, where deterministic code owns the loop and agents are bounded nodes inside it.

The deciding question for this category is who owns the loop: a factory puts phase sequencing, retries, and acceptance in code, and an agent owns only the work inside one bounded phase.

Legend: ✓ supported, ✗ not supported, ~ partial or conditional, ? not verified. Each column links to the full research note; every cell below traces to a source cited there or in the references.

The matrix
#

Feature Fabro Fluent Fusion HAR Loki Mode Machinist Ouroboros Super Simple Software Factory SuperPlane Zeroshot
Delivery single Rust binary server (Homebrew, install script, or Docker), Graphviz DOT workflow files, web UI and REST API Rust binary + skill for Codex/Claude Code/Pi, macOS only npm package (npx runfusion.ai), Homebrew formula and install script, desktop and mobile apps, web, Docker and headless; docs in-repo CLI + MCP server, npm package CLI (npm, Bun, Homebrew, Docker) + MCP server, by Autonomi Go binary and local web app with GitHub App onboarding, machinist.sh docs, optional gVisor Kubernetes runner target PyPI package + MCP server + Claude Code plugin one skill (.claude/skills/sssf), stamped into any repo self-hosted platform (Go engine, React surface), cloud or on-prem native CLI binary via npm plus one skill per harness, GraphSpec and RuntimePlan files, local, Docker target, or Cloud
Loop owner the server owns the authored DOT graph: stages, branches, loops, and human gates; agents and shell commands are nodes Rust binary owns scheduler, reviewer, tester, learner the board and engine own the loop: triage writes the PROMPT.md spec, each step flows plan, review, execute, review, and clean runs squash-merge automatically deterministic verify stages; agents own their edits the Loki engine owns contract-to-receipt; agents build and review inside it the Go controller owns invocation and recording, and the app’s standard pipeline sequences triage, optional spec, build, and independent review, while new host runs default to a native primary agent that discovers repository-defined children under advisory child reviews; agents own the work Python orchestrator owns interview, spec, execute, evaluate, evolve Python ADW scripts (code owns sequencing, retries, acceptance) the platform owns the work-order line: intake, automation stages, retries, outcome the graph engine owns sequence, parallel reviews, retry caps, and exit conditions; worker and reviewer agents are bounded nodes
Agent boundary nodes are agents, shell commands, or human gates; approval hexagons pause runs; steering reaches agents mid-turn Work Items in isolated worktrees, human attention queue, v0.3.0 Slack planning and approvals agents work stepwise inside one task’s worktree; oversight levels (off, observe, steer, autonomous) put human confirmation on merge, PR, and destructive actions by default; steering reaches agents mid-task one isolated slot per agent briefs in; review council of specialist reviewers; blocked runs ask one question named commands over named repositories, never arbitrary shell; the app adds spec-approval and human-merge gates worker never sees the grading command or expected result one agent, one named phase at a time work orders under workflow-level guardrails; ambiguous work and approvals stay with humans workers edit the checkout; reviewers return comments and never edit; the implementing agent never approves its own work
Context across seams the DOT graph plus per-node prompts; natural-language specs regenerate conforming implementations Brief, Behavior Specs (EARS), Technical Approach, Plan PROMPT.md spec with acceptance criteria per task; per-step diffs and reviewer notes on the task card; qmd markdown memory shared across sessions and nodes .har contract, shared by all agents delivery contract (the acceptance criteria) derived before any build prompts on stdin, optional prompt files immutable Seed spec frozen before any code typed JSON envelopes, parsed against a schema work-order record from intake to outcome, with event history and artifacts GraphSpec typed state plus RuntimePlan bindings; run input validated against the graph before execution
Acceptance layered verifications (test suites, linters, type checkers, LLM-as-judge) whose failures trigger fix loops deterministic final Tester bound to the reviewed commit per-step review plus configurable pre-merge and post-merge quality gates (docs, QA, security, performance, accessibility); a test gate is enforced before merge deterministic verify with per-commit evidence trail signed Evidence Receipt whose VERIFIED verdict requires one agent-independent proof by default; loki verify re-checks offline hands back a pull request, human gate, never ships (opt-in repository maintenance auto-merges only qualifying, independently reviewed docs PRs) 3-stage gate: mechanical (70% coverage default), semantic LLM (0.8), consensus on triggers gates that verify artifacts and tests after the fact result checked, failures sent back with actionable context; humans review the PR independent acceptance and code reviewers must both accept; required checks re-run after every repair; retry cap bounds the loop
Failure handling fix loops on verification failure; runs resumable from git checkpoints resumable Writer correction, fail-closed evidence, v0.3.0 recovery for long-running scheduler-owned work steps loop revise, approve, next step, rethink until done; conflicts resolved by agents in isolation before main; dependent tasks auto-rebase in order per-slot teardown, evidence kept outcomes and exit codes; an underivable contract blocks and asks one question exit code decides; a killed script restarts from the beginning; app-pipeline review findings return to the same branch for up to three repair rounds budgeted evolution, 30-generation cap, stagnation detection same-session correction, no cold restart durable runs survive retries with event history, state, and artifacts rejected work goes to a dedicated repair worker with both reviews’ feedback, then both reviews again
Self-improvement ✗ none (graphs evolve by hand in version control) ✓ Learner writes reusable Expertise after each change ✓ agents reflect after each task and update their AGENTS.md prompts and skills ~ plugins and templates reduce drift, no learner ~ memory with proof (v11.3.1) learns lessons from merged PR review comments; no per-run learner ✗ none ✓ refines specs and accumulates reusable assets across generations ✗ none ~ continuously re-evaluates which issues agents can handle; no learner ✗ none (profiles reuse setups, nothing learns from outcomes)
Isolation ~ optional Daytona cloud sandboxes with network and filesystem isolation; laptop runs local ✓ isolated worktree, remote via AWS Fargate ✓ branch and worktree per task; touched files tracked per task with auto-rebase of dependents ✓ worktree, ports, and database per slot ~ single-run isolation undocumented; –attempts N runs N scored worktree attempts (v11.3.1) ~ optional gVisor Kubernetes runner target (one Job per attempt, restricted egress, no GitHub credentials to the agent, Codex-only at launch); local runs executor-dependent ~ delegated to the host runtime ✗ runs on current branch, no sandbox or merge step ? not verified in the fetched docs ~ local runs edit the current worktree unisolated; the Docker target moves execution to your infrastructure
Coding agent any agent CLI or shell command as a node; models routed per node (Claude, Gemini, GPT, Kimi named) Codex, Claude Code, or Pi Anthropic, OpenAI, Gemini, Z.ai, Ollama, plus Hermes, Droid, Cursor, and any ACP agent; five model lanes (executor, planner, validator, title, workflow); built on Pi Claude Code, Cursor, Codex, or any MCP agent Claude (default); Cline, Codex, Aider, or OpenCode experimental (your keys) any executable reading a prompt on stdin (Kubernetes runner target: Codex only at launch) 14 runtimes, Claude Code through Antigravity pi only (claude_code stubbed) Claude, Cursor, OpenAI, OpenRouter, Perplexity via integrations Codex, Claude Code, GitHub Copilot CLI, or Pi as worker and reviewers (your logins)
Trace durable event stream, checkpoints, conclusions; SQL via DuckDB; web UI Work Item / Attempt record with bound evidence Command Center: tokens, USD cost, autonomy ratio, per-agent output, model mix, eleven views with CSV and OpenTelemetry export; per-task run history Mission Control dashboard per run and artifact Evidence Receipt plus verdict and cost; live control plane on the site durable events, artifacts, outcome, duration, reported tokens event sourcing with full replay and lineage SQLite, tool calls visible mid-run Run records per step: inputs, outputs, retries, cost durable SQLite ledger of node results; runs UI; restart or resume failed runs
License MIT Apache-2.0 MIT Apache-2.0 BUSL-1.1 (source-available) MIT MIT MIT Apache-2.0 core, Enterprise Edition license for /ee MIT
Pricing free, MIT, self-hosted; one-click Railway deploy; no hosted tier free, no paid tier, no hosted cloud free, MIT, self-hosted on your nodes; no hosted tier; early preview core free; HAR HQ Team $400 per month up to 50 users with $100 monthly cloud credits, Enterprise custom (self-hosted or VPC, SSO/SAML/SCIM) free source-available CLI, your keys; no published hosted prices as of 2026-10-09 free, early access free, GitHub Sponsors funded, enterprise offerings conditional free, self-hosted free open-source engine to self-host; Business $199 per organization per month with $50 shared monthly usage, OSS Projects free, Enterprise custom (as of 2026-10-10) free MIT core on your harness logins; Zeroshot Cloud managed target without published prices and unadvertised since 2026-10-19
Born 2026-03-13 2026-07-10 2026-04-13 2026-06-28 2025-12-26 2026-07-16 2026-01-14 2026-08-02 2025-05-07 2025-12-25
Stars about 1.7k about 110 about 1.3k about 120 about 1.1k about 500 about 6.2k about 950 about 7.7k about 2.0k

Reading the matrix
#

The loop-owner row is the whole category in one glance: SSSF and Fluent move the loop into code (Python ADWs, a Rust scheduler), SuperPlane’s platform and Loki Mode’s engine own it as a product, Fusion’s board and engine own it step by step (plan, review, execute, review per task step, auto-merge when the gates pass), while HAR coordinates a fleet but leaves each agent free to edit, which is why its acceptance is verify-focused rather than loop-owned. A factory is defined by determinism, and HAR earns its place through deterministic verification rather than deterministic sequencing. Fabro and Zeroshot author the loop as a graph: Fabro’s DOT files make the process a version-controlled artifact with human gates, and Zeroshot’s built-in review loop adds the acceptance rule the category keeps circling, the agent that wrote the code never approves it.

The isolation row is where the category separates from a single long agent call: Fluent and HAR isolate every candidate in its own worktree, Fusion isolates every task in its own worktree and auto-rebases dependent tasks in merge order, SSSF deliberately runs on the current branch with no sandbox or merge step, the gap its own note names first, Loki Mode documents worktrees only for its scored –attempts N runs, and SuperPlane’s isolation stays unverified in the fetched docs. Fabro’s sandboxes are optional cloud VMs and Zeroshot’s local runs edit the current worktree, both partial by default.

The self-improvement row is now three learners and a partial: Fluent writes reusable Expertise after accepted changes, Ouroboros refines specs and accumulates assets across a budgeted evolution loop, Fusion’s agents rewrite their own AGENTS.md prompts and skills after each task, and Loki Mode’s memory with proof learns lessons from merged PR review comments, while SSSF, HAR, Machinist, Fabro, and Zeroshot stay static and SuperPlane only re-scores its intake. The difference is what compounds: Fluent compounds lessons about the code, Ouroboros compounds the specification itself, Fusion compounds the agent instructions, which is the least inspectable of the three because the rewritten prompt is the input to the next run.

The hidden-grading row is Ouroboros’ claim to a distinct cell: it is the only member that structurally withholds the grading command and expected result from the worker agent, an anti-reward-hacking design none of the others attempt. Zeroshot’s independent reviewers are the closest neighbor, but the reviews are visible LLM judgments rather than hidden grading commands.

The narrow columns are candid: SSSF’s pi-only dependency versus the open agent matrix of Fluent and HAR is the single biggest reason to look past the founding member’s headline ease. Ouroboros runs the other direction with 14 runtimes, the widest agent boundary in the category, paid for in evaluation tokens and beta churn. Machinist narrows the boundary in the other direction: its named-command entrypoint stays the category’s strictest agent interface, paid for in restart-from-zero script semantics, with isolation now opt-in in a Codex-only gVisor Kubernetes target. The two graph-authored newcomers sit between the poles: Fabro’s authorable DOT graphs are the widest process surface, paid for in nightly-only releases, and Zeroshot’s fixed review loop is the tightest acceptance story, paid for in a closed contribution door and a four-harness ceiling.

The pattern has independent practitioner validation beyond this category’s tools. Will Larson’s Imprint runs a Linear-project factory loop in production and finds it compounds only when the surrounding pieces exist: one source of task state, metric access, and a harness that works without a laptop. The term’s AI-era origin traces, by Larson’s own attribution, to Justin McCarthy’s February 2026 StrongDM essay, which describes the same bet from the inside: specs and scenarios drive agents, and no human writes or reviews the code.

The rejected long tail: agentic-software-factory (1 star, inflated agent counts), ai-factory (4 stars, no license), and Takk8IS/software-factory (0 stars, no license) did not clear the citation or credibility bar and are named here so no future run re-adds them silently. The 2026-10-07 entrant scan resolved the topic’s previously untracked candidates too: the AgentField family (SWE-AF at 1,030 stars and CodeAF at 296 stars, one in public beta and one in early preview on the AgentField platform) stays out on a scope call for the owner, whether a platform family gets a column, mirroring retrieval’s Haystack precedent; Finn-loop (319 stars, dormant since 2026-07-23) is a three-skill Linear recipe rather than a loop-owning factory, addyosmani/factory (222 stars) is a dormant reference demo, mitkox/esf (187 stars) is an 18-day-old fork of this category’s own Machinist, and lago-morph/software-factory (1 star, no license, dormant since 2026-06-08) closes the tail. The 2026-10-10 sweep of the subject’s first dedicated list (awesome-software-factories, 17 stars) resolved the rest: langchain-ai/open-swe (10.9k stars) is flagged to the harnesses category because its LLM graph owns plan, implement, and validate rather than a stamped pipeline, strongdm/attractor (1,328 stars) is StrongDM’s dormant nlspec artifact rather than a runnable member, Runfusion/Fusion (1,261 stars) and jonwiggins/optio (1,062 stars) are active orchestrators with no independent discussion found, vercel-labs/eve-software-factory-template (1,160 stars) and mastra-ai/softwarefactory-template (54 stars) are platform templates, dzhng/skills (1,022 stars) is a skills pack, bastani-inc/atomic (855 stars) ships no standard license, thruwire/foreman (731 stars) holds its coverage-bar rejection, and the sub-400 tail (Genie, taskplane, Anneal, DoorDash’s agentic-orchestrator, great_cto) stays below the admitted cohort’s footprint. The 2026-10-11 cross-check of the same list (fetched again 2026-10-11) resolved its remaining untracked entries. garrytan/gstack (135,878 stars as of 2026-10-11, the largest repository any list in this space names) brands itself an open source software factory, but it is a slash-command skill pack whose sequencing belongs to the human running /office-hours, /review, and /ship, so it is flagged to skills rather than admitted here. gastownhall/gascity (1,360 stars, MIT, active) is Gas Town’s reusable infrastructure repackaged as an orchestration-builder SDK, flagged to orchestration beside its parent note. Leeroo-AI/kapso (121 stars, MIT, active) runs long-horizon optimization campaigns over AI and data systems with published MLE-Bench, ALE-Bench, RelBench, and IOAI results, flagged to automated research beside DeepAnalyze. CodeMachine-CLI (2,510 stars) has been dormant since 2026-02-25 and the list itself marks it unmaintained. VisionForge-OU/foreman (443 stars, no standard license, dormant since 2026-06-29) is a different repository from the already-rejected thruwire/foreman. coleam00/ai-software-factory (365 stars, active, no license) is a setup template whose factory logic lives in Archon’s SDLC pack, the same platform-template pattern as the rejected eve and Mastra templates. Orbi (213 stars, AGPL-3.0, active, label-an-issue-get-a-merged-PR with independent review) has zero independent discussion and stays below the admitted cohort’s footprint. ramure (224, no license, dormant since 2026-06-30), squid (203), inkwell (161, the member SSSF’s sandbox companion, dormant since 2026-08-09), dark-factory-experiment (168, no license), claude-lane-stack (122), sgai (137, no standard license), and the sub-100 remainder close the tail. The list’s 2026-10-12 weekly sweep added sixteen entries, four of them already resolved (tuios and Markus by their 2026-10-10 rejections, desplega-ai’s agent-swarm and GraphCode as orchestration members), and the other twelve are parallel-workspace surfaces and terminal multiplexers rather than loop-owning factories (Synara at 2,099 stars the largest, then wmux, Calyx, Aizen, Intent, Roamgate, Rovai, Rove, CliDeck, Dray, tlbx, and ReevesAgents), flagged to the surfaces and orchestration workers rather than admitted here. Helmor (1,309 stars, Apache-2.0, dormant since 2026-08-22), an earlier list entry the prior sweeps never named, is a local-first workbench for orchestrating coding agents, a worktree-per-task workbench the human drives with no phase-owning engine of its own, flagged to orchestration beside its peers rather than admitted here. The sweep’s factory-section additions resolved below the bar: patchmill (15 stars), smile-software-factory (53, no license), nicolasmelo1’s software-factory (34), stratonext’s software-factory (11), and sjarmak’s reliability-patterns companion repository (8). The list’s 2026-10-15 weekly sweep added nine tools, one benchmark, and three field reports, and its one factory-shaped candidate stays below the bar: uzi (vtmocanu/uzi, 91 stars, MIT, alpha, an issue-label dark factory with plan approval and human merge) fails on the Orbi precedent (zero independent discussion, a self-published write-up as its only coverage, a nightly RC train with no stable release), with a re-review trigger on independent coverage or a stable release. The sweep’s other additions routed to their kinds: shipyard, yume, and SuperAgent to the parallel-workspace surface class, agentbox and pi pod to sandboxing, aislop to code review, OpenCode++ to harnesses, oz-agent-action folded into the Warp Factories sibling note, Autoheal held on the standing closed-commercial line, CRBugBench flagged to evaluation review, and the field reports to people and publications. The list’s 2026-10-19 weekly sweep added two tools and a field report, and its one factory-shaped candidate stays below the bar: Antfarm (snarktank/antfarm, 2,495 stars, MIT, dormant since 2026-02-26) installs a fixed specialist-agent team into OpenClaw rather than owning a loop, and fails on the CodeMachine-CLI dormancy precedent; gwq (a git-worktree fuzzy finder) and the loom-to-Overalls rename (a Firecracker microVM dev-machine platform) are isolation-class, flagged to the sandboxing worker, and the WarpBench field report (Warp’s cost-per-PR case study) to people and publications. The list’s 2026-10-22 sweeps added five tools and three field pieces, and none are loop-owning factories: ccpm (a GitHub-issues task queue with one agent per issue) routes to task management, takopi (a Telegram bridge with per-session queues) to surfaces, Repowire (inter-agent discovery and messaging across projects) to orchestration, parallel-worktrees (a Claude Code worktree skill) to orchestration beside Worktrunk, and Codealong (a desktop board over every local agent session) to task management, pi pod was already sandboxing-routed on 2026-10-15, and GPT-Pilot (the spec-writer-architect-developer-reviewer-debugger app builder) rejects on the CodeMachine dormancy precedent with the list’s own unmaintained tag and a disclosed malicious-code window in its README; the two AI Engineer Europe talks and the madewithlove owning-the-factory field report went to people and publications.

Choosing from the matrix
#

  • Want the easiest code-owned loop on a single agent and will wire your own gates: Super Simple Software Factory.
  • Want a self-improving factory with a real final Tester and worktree isolation, pre-1.0 accepted: Fluent.
  • Want a fleet of coding agents on one repo with deterministic verification and evidence, keeping your agents: HAR.
  • Want a local, auditable runner with a strict named-command boundary around any agent CLI, early-access accepted: Machinist.
  • Want vague briefs turned into verified code with the grading hidden from the worker, 14 runtimes and beta churn accepted: Ouroboros.
  • Want every change accepted by independent reviewer agents that can never approve their own work, on your existing Codex or Claude login: Zeroshot.
  • Want your process itself version-controlled as an authored graph with human gates and model routing, nightly releases accepted: Fabro.
  • Want an autonomous local issue-to-PR loop with an offline-checkable receipt, source-available licensing accepted: Loki Mode.
  • Want the whole intake-to-PR line owned by a platform with durable runs and confidence-gated backlog intake: SuperPlane.
  • Want a board-first factory you can steer from any node you own, with per-step gates and auto-merge on green, early-preview churn accepted: Fusion.

Changes
#

  • 2026-08-29 - Created as a single-column scaffold for SSSF when the Software factory category was seeded.
  • 2026-08-29 - Rebuilt to three columns (SSSF, Fluent, HAR) with the who-owns-the-loop thesis.
  • 2026-08-30 - Extended to four columns with Ouroboros, self-improvement and hidden-grading prose updated.
  • 2026-08-30 - Re-sorted columns alphabetically with Fluent first, dropping the founding-member-first convention.
  • 2026-09-05 - Extended from four to five columns with Machinist.
  • 2026-09-13 - Added the Pricing row, recording HAR HQ’s first published pricing (Team $400 per month up to 50 users with $100 monthly cloud credits, Enterprise custom) alongside the free open-source core.
  • 2026-09-24 - Removed the verification preamble line on owner request.
  • 2026-09-25 - Refreshed the HAR (about 95, after v1.14.3) and Machinist (about 460) star cells.
  • 2026-09-27 - Refreshed the SSSF star cell (about 900).
  • 2026-09-29 - Refreshed the Machinist star cell (about 470); all other cells re-verified unchanged.
  • 2026-10-02 - Refreshed the star cells (Fluent about 110, Machinist about 480, Ouroboros about 6.2k, SSSF about 920); HAR and all other cells re-verified unchanged.
  • 2026-10-06 - Added the independent-practitioner paragraph grounding the loop-owner thesis (Will Larson’s production factory loop, Justin McCarthy’s term-origin essay) with two references; refreshed the Machinist and SSSF star cells.
  • 2026-10-07 - Extended from five to seven columns with SuperPlane and Loki Mode, re-sorted alphabetically; recorded the 2026-10-07 entrant-scan resolutions in the long-tail paragraph; all previous cells re-verified unchanged.
  • 2026-10-09 - Moved three Loki Mode cells on the v11.3.1-v11.3.9 release evidence: self-improvement from none to partial (memory with proof learns lessons from merged PR review comments), isolation from unverified to partial (–attempts N scored worktree attempts), and the coding-agent roster updated to Claude default with Cline, Codex, Aider, and OpenCode experimental; all other cells, including HAR HQ pricing and the star row, re-verified unchanged.
  • 2026-10-10 - Extended from seven to nine columns with Fabro and Zeroshot, re-sorted alphabetically; moved the SuperPlane pricing cell to the website’s first published plans (Business $199 per organization per month with $50 shared monthly usage); recorded the awesome-software-factories scan resolutions in the long-tail paragraph; graph-authored-loop, isolation, self-improvement, and hidden-grading prose updated for the new columns.
  • 2026-10-11 - Refreshed the HAR (about 110) and Machinist (about 500) star cells and recorded the 2026-10-11 awesome-list cross-check resolutions in the long-tail paragraph (gstack, Gas City, and Kapso flagged to skills, orchestration, and automated research; CodeMachine CLI, VisionForge’s foreman, the Cole Medin template, Orbi, and the small tail rejected); all other cells re-verified unchanged.
  • 2026-10-11 - Extended from nine to ten columns with Fusion (the MIT board-first factory built on Pi, with per-step plan-review-execute-review gates, worktree auto-rebase, and multi-node mesh), inserted alphabetically between Fluent and HAR, every row gaining a cell traced to the new note; this reverses the 2026-10-10 long-tail holding on the orchestration resolution’s flag and the fuller mechanism evaluation (the coverage caution stays in the note); the loop-owner, isolation, and self-improvement readings updated, and the choosing section gained a Fusion bullet.
  • 2026-10-13 - Refreshed the HAR (about 120) and Zeroshot (about 2.0k) star cells; recorded the awesome-software-factories 2026-10-12 weekly-sweep resolutions in the long-tail paragraph (twelve new parallel-workspace entries flagged cross-category, Helmor flagged to orchestration as an unresolved prior entry, the factory-section additions rejected below the bar); all other cells re-verified unchanged.
  • 2026-10-16 - Moved Machinist’s isolation cell from unsupported to partial (the gVisor Kubernetes runner target, Codex-only at launch) and extended its delivery, loop-owner, agent-boundary, failure-handling, and coding-agent cells for the 2026-10-15 app release; the boundary prose updated; recorded the awesome-software-factories 2026-10-15 weekly-sweep resolutions in the long-tail paragraph (uzi below the bar on the Orbi precedent, the rest routed to their kinds).
  • 2026-10-17 - Extended Loki Mode’s acceptance cell for the seal VERIFIED default requiring one agent-independent proof, from the v11.17.33 release notes; all other cells re-verified unchanged (Fabro nightly train, Fusion and Ouroboros release trains, SuperPlane v1.0.0-rc1, HAR HQ pricing).
  • 2026-10-21 - Moved Zeroshot’s coding-agent cell to four harnesses (Pi added in v10.14.0) and its pricing cell to the v10.15.0 README de-advertising Zeroshot Cloud; the boundary prose’s three-harness ceiling updated to four; recorded the awesome-software-factories 2026-10-19 weekly-sweep resolutions in the long-tail paragraph (Antfarm rejected on the dormancy precedent, gwq and the Overalls rename flagged to sandboxing, the WarpBench field report to people and publications); all other cells re-verified unchanged (Fabro nightly train, Loki v11.18.0, Ouroboros v0.59.3, SuperPlane rc2/rc3, Machinist commit count, HAR HQ pricing).
  • 2026-10-23 - Extended Machinist’s loop-owner cell for the native-agents default run mode (advisory child reviews) and its acceptance cell for the opt-in repository-maintenance auto-merge limited to qualifying docs PRs; recorded the awesome-software-factories 2026-10-22 sweep resolutions in the long-tail paragraph (five tools routed to task management, surfaces, and orchestration, GPT-Pilot rejected on the dormancy precedent, the field pieces to people and publications); all other cells re-verified unchanged (Fabro nightly train at v0.393.0-nightly.1, Loki v11.60.x burst, Ouroboros v0.59.7, SuperPlane rc5 with runner v0.2.1, Fusion beta.9 with npm latest still 0.77.0, Zeroshot v10.16.0, HAR HQ pricing).

See also
#

References
#