Building an (almost) fully self-hosted, sandboxed, agentic software factory

The strongest takeaway was that code generation is not the bottleneck anymore. Verification is. Several people said unit, integration, and end-to-end tests written by agents still create a human review burden, because you are often checking machine-generated code with machine-generated assumptions. The more credible pattern was to build a hermetic environment that mirrors production, drive it from the outside through browsers, terminals, and APIs, and make the results reviewable through recordings and logs. That turns verification from reading piles of generated tests into judging observable behavior. Others pushed similar ideas like cross-model review, parallel candidate solutions, and simulation-based tests that are designed to fail against known buggy variants. The GPU subthread landed in a pragmatic place. Local models are already useful for cheap support work like transcription, embeddings, browser actions, image analysis, and lower-stakes coding tasks. Multiple people reported decent results with Qwen on consumer hardware like a 3090 or 5090. But the same people also said local models still lag frontier systems on hard agentic coding. They think too long, run out of context, loop on discovery, or become too slow to stay in flow. For now, local inference looks more like a cost-saving sidecar than a full replacement. There was also a sharp undercurrent of skepticism about the "walk away while it ships" vision. People who use these tools daily said they work great for personal automation and for getting through the first 80 or 90 percent. They break down on the ugly edge cases that decide whether a product is shippable. That is especially dangerous if the human operator does not understand the generated code well enough to rescue the last mile. So the conversation ended up less excited about dark factories and more interested in the missing control plane around them: sandboxing, reproducibility, behavioral verification, and a sane human merge point.