Ask HN: What are you working on? (August 2026)

The strongest pattern was infrastructure for agentic development itself. There were multiple attempts at cloud devboxes, harnesses, orchestration layers, memory systems, safety guards, sandboxing frameworks, and workflow engines for long-running coding agents. Several projects framed the same pain point in slightly different ways: coding agents are useful, but only if you can constrain them, preserve context, inspect failures, and hand work across tools without losing the plot. That turned into products for session recall, VM isolation, command guards, task planning, and persistent cloud environments. The conversation has clearly moved past “can an agent write code” and into “what scaffolding makes agents dependable enough to own part of the software lifecycle.” A second pattern was local-first or self-hosted software meeting AI. Many builders want the convenience of AI features without giving up data control, recurring cloud costs, or portability. That showed up in offline meeting document generation, end-to-end encrypted workspaces, personal data copilots, local photo cataloging, browser-based model inference, Kobo and Pebble hacks, and products explicitly built around bring-your-own-model or bring-your-own-key setups. The consistent framing was practical, not ideological. Builders want AI where it helps, but they do not want their whole product economics or trust model tied to hosting inference themselves. There were also several unusually concrete vertical products, which is often where these threads produce the most real startup signal. The most enthusiastic response went to Sawdust, a woodworking design and planning tool that models real lumber, joinery, shop tools, and bills of materials instead of generic CAD primitives. Preloop, which runs unmodified GitHub Actions locally or self-hosted inside microVMs, also drew strong interest because it attacks a clear pain point with a technical wedge. Other serious vertical efforts included an automotive diagnostic app grounded in OBD2 data plus make-specific interpretation, a transit software managed service around OneBusAway, a congressional-hearings search product, a product-photo editor for online sellers, a funeral-home ops platform, a pinball score scanner, and scientific or academic information systems built by domain insiders. Games and creative tools were still a major share of the thread, but even there the tone was different. Many creators said outright that AI removed enough implementation friction to let them pursue old ambitions like niche game engines, homebrew console tooling, ports, simulators, puzzle games, and media apps. The key change is not that AI made the ideas better. It made many projects cheap enough to exist at all. That came through repeatedly from people reviving abandoned ideas, tackling rewrites, or finally shipping things they had postponed for years. The general mood was upbeat, generous, and unusually supportive. People traded feedback, bug reports, feature ideas, packaging advice, and market suggestions with little cynicism. The comments that landed best were attached to tools with obvious lived pain behind them. Products built from real workflows, like woodworking, CI debugging, transit ops, diagnostics, or private knowledge management, got far more traction than generic “AI for X” pitches. The clearest takeaway from the whole thread is that the new bottleneck is no longer raw implementation. It is taste, workflow fit, trust, distribution, and the discipline to choose a narrow problem worth solving.