My first software factory was arguably the humble Ralph loop, 14 months ago. Then I wrote IssueOps, a factory that ran entirely in GitHub Issues and Actions, so I could offload my laptop and keep building while travelling. That cost too much. API billing, amirite? So I brought the build farm home and Anneal was born. It ran on my Claude subscription locally and wrapped Ralph loops with outer loops, guardrails and tickets. But it lived inside each project, and every copy soon drifted. About 12 weeks ago I started Holodeck. Instead of putting the factory in the project, I build the projects in the factory. The first project it built was Holodeck itself. What I've learned: 1. Deterministic gates are the real quality mechanism. The model isn't trusted to judge its own work. Hooks and tests run outside the agent loop. Cheaper models fail and retry instead of committing bad code. 2. The same agent never writes and grades its own work. Test author, implementer and reviewer are separate agents, on different models. 3. Detailed documents first, fast execution after. The spec is now the hard part. 4. Fresh contexts. Always. Tickets keep a log of what was tried and failed so the next loop doesn't make the same mistake. 5. An orchestrator with subagents. The top model coordinates and doesn't do grunt work. 6. Isolation by default. Every agent that edits code gets its own worktree. Runners are throwaway Docker containers with no Docker socket, no source mount and no host ports. 7. Agents get references to secrets, never the values. Enforcement sits in the tooling and hooks, not in the prompt. (I have a product for this coming.) 8. Store a detailed event log. Agents can interact with the API to emulate usage, see real-time logs, and self-improve. Even self-heal. 9. Tools are disposable. Every model release makes some tools redundant, so refactor your stack.

I love this, but I wonder how it works under the hood. @AnthropicAI models assume that ALL CAPS INSTRUCTIONS are emphasised or more important. Other providers, particularly @OpenAI models consider it RUDE SHOUTING 😆 So if Claude reads AGENTS.md, does it read it like it would read CLAUDE.md, or like other harnesses read AGENTS.md? 🤔 The mod code looks like there's no translation layer, so will AGENTS.md now need to be written for the lowest common denominator? Or we'll still need to keep CLAUDE.md for now too, I guess? /cc @adocomplete

Searching Reddit has been the most annoying thing for me. I wished I had semantic search to surface posts I'm interested in. So I built ThreadScope. I use it to find posts relevant to my interests (IoT, home automation, self-modifying assistants) and to the dev community I serve (developer security, AI safety). Another user uses it to stay on top of current trends. Another works at a marketing company and uses it for competitor analysis and reputation control. I didn't see that use case coming, although it seems obvious now. Need something like this in your life? Try the seven-day free trial 😁 threadscope.io/?ref=x

A year ago I wired a chat box to Claude Code, handed it a Home Assistant API key, and asked for a battery dashboard. It immediately rewrote itself: navigation, a dashboard, a list of devices low on juice. It also pulled the key straight into context 🤣 That's how I accidentally started a dark software factory. Step one was convincing a self-modifying web app to harden itself. Step two was starting from scratch. Mine's called Holodeck, and it runs like the marketing agencies I worked at 20 years ago. An account manager interviews the client, a tech lead writes the spec, developers pull tickets off a Kanban board, QA throws them back. Every role is an agent. @Steve_Yegge's "Fences, not Sandboxes" got me writing this up. "Seats and Sunsets" forced me to finish it. I agree with him on seats. On fences, less so. Instead of hundreds of rules in context, I use a handful of deterministic gates outside it. Nothing merges unless they all pass, and the agents never read the rules. A weaker model behind a real gate can't ship a bad merge. It can only cost me retries. First in a series. What should I dig into next: the state machine, the gates, or the bit where I still poke agents with a stick? bendechr.ai/articles/2026-09…

How are developers actually using LLMs? I polled my audience at @kc_dc last week, arguably the most curious developers in the US Midwest, and the numbers might surprise you. Have you used an LLM, at work or personally: - for anything related to writing code: 100% - to autocomplete a function: 50% - to build non-stop for 5 minutes: 5% - to take a spec and run for more than 4 hours: 3% - for more than 12 hours: 2% - to build your own dark software factory: 1% Half the room has never had an LLM autocomplete a function. 95% have never let one run unattended for five minutes. So if you feel like you'll never catch up, you're in the majority. That's worth remembering as a community. There's so much noise about what people are doing with LLMs that it's easy to think you're falling behind, when most of us are still near the starting line. As I told the room: be curious, play, take it one step at a time. You'll catch up faster than you expect. Where are you on this journey? And if you're in the 5%, how are you helping others get there?

Claude is creating poor working conditions in my factories and all the workers have gone on strike 😆

Watch piped.video/ErMSHiQRnc8 and tell me you weren't willing momentum on 😆

People keep telling me dark software factories aren't possible. I continue to disagree. And now, it seems, so does @tessl_io, where Maria had Kikimora build itself ♺ That's exactly how I built mine too. The concept that we don't need to look at the code anymore is contentious, and I get it. It's another reminder that coding will soon be a hobby, not a career. But it's already the case. There's no going back. The hard part now is turning user requirements into technical specifications. That's where software engineers and solutions architects still have value, right? Well, when, and not if, that is automated too, what then? That's what I'm building as a proof of concept. As much as I hope it won't work, my hypothesis is it will. Because everything we thought wouldn't be working two years ago is working today. I have a blog post in my head about this, especially around the work involved in making this something someone who's not already an engineer can use. What about you? Have you built something like this? What was the hardest part for you?

Every GTM team we talk to is trying to automate something right now. Very few of them have a clear answer for where automation should stop. We sat down with developer relations leader @bendechrai to talk through exactly that. Where the line sits, what he automates without a second thought, and what he refuses to hand over to AI no matter how much time it would save him. GTM Engineering and how to build AI harnesses is out later today on our YouTube channel. Keep an eye out for it.

@bendechrai, @LaraNerdsom & @brentestewart present on #Frontend Development this July at Nebraska.Code(). nebraskacode.amegala.com/ #SoftwareCraftsmanship #WebApp #AppSecurity #Angular #API #Blazor #BlazorComponents #DOTNET @Cisco @SourceAllies @AlienArcTech @AlloFiber #TechConf

@bendechrai and @alignedDev present on #SoftwareCraftsmanship this July at Nebraska.Code(). nebraskacode.amegala.com/ #FrontEnd #AppSecurity #SpecKit #AI #TechConf #HeartlandDev @TigerpawCo @DataShieldCorp @CultivationCorr @CushWakeUS @digitalskyco @discover_dynamo @DistrictOperat1