--- title: 3,000 commits a day: the OpenClaw 'dark factory'... description: A roundup of practitioners running 10-20 parallel AI coding agents simultaneously — described as 'dark factory' workflows — reveals both the infrastructure... canonical: https://daily.dev/posts/3-000-commits-a-day-the-openclaw-dark-factory-workflow-is-wild-and-it-might-actually-work-s9l35owp8 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: 3,000 commits a day: the OpenClaw 'dark factory' workflow is wild, and it might actually work | daily.dev og:description: A roundup of practitioners running 10-20 parallel AI coding agents simultaneously — described as 'dark factory' workflows — reveals both the infrastructure... og:url: https://daily.dev/posts/3-000-commits-a-day-the-openclaw-dark-factory-workflow-is-wild-and-it-might-actually-work-s9l35owp8 og:image: https://api.daily.dev/og/posts/s9l35OwP8.png og:image:alt: 3,000 commits a day: the OpenClaw 'dark factory' workflow is wild, and it might actually work og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # 3,000 commits a day: the OpenClaw 'dark factory' workflow is wild, and it might actually work **[Trends](https://daily.dev/sources/trends)** · 4 min read · 5 upvotes · 0 comments ## Summary A roundup of practitioners running 10-20 parallel AI coding agents simultaneously — described as 'dark factory' workflows — reveals both the infrastructure challenges and philosophical tensions emerging at scale. Key failure modes include workspace loss, silent tool-call hangs, credential expiry, and container health vs. agent health confusion. A DoorDash engineer's Research-Plan-Implement framework and Molted's managed runtime checklist address reliability. Matt Pocock raises a critical concern: 'instruction rot,' where agents that update their own instructions become impossible to steer. Nvidia is positioning itself as enterprise infrastructure provider for agent fleets via NemoClaw and OpenShell. The honest takeaway is that AI amplifies work density without reducing hours, and the steerable-system problem will matter most as these setups scale. ## Content ## The setup A cluster of posts this week paints a picture of where the serious AI-coding crowd has landed: not using Copilot to autocomplete a line here and there, but running fleets of autonomous agents around the clock, acting less like developers and more like factory floor managers. Vincent Koc, a core maintainer of OpenClaw, describes his workflow as a "dark factory" — 10 to 20 parallel AI coding agent "swim lanes" running simultaneously, generating up to 3,000 commits per day. He organizes agents by task type (CI, features, bugs), uses Git worktrees, and builds reusable "skills" files so agents can pick up context without hand-holding. His framing is deliberate: he's not writing code, he's managing a production line. Steipete is running something similar. His setup wakes up every five minutes, reviews open threads, and routes work to agents. When someone files an issue on one of his open source projects, an agent called Claw Sweeper reviews it against a `VISION.md` file, and if it fits, creates and auto-reviews a PR. The whole thing runs without him in the loop for most decisions. Peter Yang has taken it further in a different direction — he now writes his social media posts through Codex, chaining browser use and skill composition to handle the whole workflow. ## The infrastructure problem nobody talks about Running agents locally is easy. Running them reliably at 3am when nobody's watching is a different problem entirely. A post from the team behind Molted (a managed runtime for agent fleets, disclosed upfront) lays out the failure modes that don't show up in demos: workspace loss on restart, silent tool-call hangs, browser session drift, memory spikes, credential expiry. Container uptime, they point out, is not the same as agent health. A container can be running fine while the agent inside it has been stuck waiting on a hung tool call for six hours. The checklist they propose — persistent workspaces, meaningful restart semantics, per-agent resource limits, scoped secret handling, human override paths — reads less like a product pitch and more like a list of things people have already gotten burned by. A DoorDash engineer's writeup covers similar ground from the practitioner side. He tried multi-agent teams first, got inconsistent output because agents lacked shared context, then rebuilt around a Research-Plan-Implement framework with git worktree isolation and crash recovery. His tool, Agentic Orchestrator, drives features through deterministic phases so the whole thing doesn't collapse when one step goes sideways. He's also honest about something most of these posts skip: the cognitive load is brutal. AI amplifies work density, he says. It doesn't reduce your hours. ## The pushback Not everyone is enthusiastic about the self-improving, self-steering direction this is heading. Matt Pocock put it plainly: "I have a deep distrust of almost any 'self-improvement' loop in coding agents." He's talking about automatically created memories, CLAUDE.md suggestions applied after every session, agents that update their own instructions. His concern isn't that the suggestions are always wrong — sometimes they're fine. The problem is that agents over-index on them in ways that make the whole system impossible to steer. Each project ends up "unsteerable in its own way." He's calling it "instruction rot" and asking if there's a better name. It's a real tension. The dark factory crowd is building toward more autonomy, more self-direction, more agents writing their own `AGENTS.md` files. Pocock is pointing at the failure mode that comes after that: a system that's technically running but that nobody can actually redirect. ## Where Nvidia fits in Nader Khalil, Nvidia's Director of Developer Technologies, offered a deliberately plain definition of what an agent actually is: "an LLM and a harness." Nvidia has full-time contributors on OpenClaw and is working through its PR backlog. Their enterprise play is NemoClaw — blueprints for companies that want to adopt agent harnesses without building the infrastructure from scratch. OpenShell is their security runtime layer for enterprise deployments. Khalil's prediction: every enterprise will eventually build specialized AI agents. Nvidia is positioning itself as the picks-and-shovels provider for that transition, with partnerships at CrowdStrike, Cadence, and Palantir already in place. ## The honest middle ground One post worth flagging cuts through the hype in a useful way: AI code review doesn't need to be perfect to be worth doing. Catching half the issues before they reach a human reviewer is enough to shift whether a team keeps up or falls behind. That's a more defensible claim than "AI replaces reviewers" — and it's probably closer to what's actually happening in most of these workflows, even the aggressive ones. The dark factory is real. The question Pocock is raising — whether the thing you've built is still steerable — is the one that's going to matter most as these setups scale. --- Tags: [#cicd](https://daily.dev/tags/cicd), [#ai-agents](https://daily.dev/tags/ai-agents), [#codex](https://daily.dev/tags/codex) [View this post on daily.dev](https://daily.dev/posts/3-000-commits-a-day-the-openclaw-dark-factory-workflow-is-wild-and-it-might-actually-work-s9l35owp8) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"3,000 commits a day: the OpenClaw 'dark factory' workflow is wild, and it might actually work","url":"https://daily.dev/posts/3-000-commits-a-day-the-openclaw-dark-factory-workflow-is-wild-and-it-might-actually-work-s9l35owp8","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/3-000-commits-a-day-the-openclaw-dark-factory-workflow-is-wild-and-it-might-actually-work-s9l35owp8"},"datePublished":"2026-06-05T16:27:02.620Z","dateModified":"2026-06-23T09:26:58.762Z","description":"A roundup of practitioners running 10-20 parallel AI coding agents simultaneously — described as 'dark factory' workflows — reveals both the infrastructure...","image":"https://i.ytimg.com/vi/pmoDeA3RBZY/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/pmoDeA3RBZY/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/3-000-commits-a-day-the-openclaw-dark-factory-workflow-is-wild-and-it-might-actually-work-s9l35owp8","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"cicd,ai-agents,codex","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"3,000 commits a day: the OpenClaw 'dark factory' workflow is wild, and it might actually work"}]} ```