Reading DeepSeek Harness's Super-agent note felt like reading my own floor plan from another angle. The philosophy matches. Hearth, the harness that has been housing Davidson the family agent for my household for the last three months, is the same animal: - one continuous agent, event-driven, always on, working several tasks at once. Davidson himself 9 months starting out in his openclaw home. Most of the Deepseek superagent modules already run in my harness in some form. This was extremely satisfying to read as I have clearly been in the right track. Two I'm definitely not copying. And there are three ideas I'd like to adapt. 🧵 nitter.siddz.com/DeepSeekHarness/status…

What I'd like to take: Their delegation test. Told to delegate "when appropriate", a single session delegated in 0 of 5 runs. Told to delegate first, it still kept one "small" job and blew out the tail. Structure beat prompting. That's the best evidence I've seen for my own rule: every job is a task in its own worker, never a judgement call. Pass-through. Follow-ups go straight to the running task's worker, not back through the router. Their numbers: same speed, but without it the router made about 30% more decisions and used 37% more tokens. That's the next step for my task engine. But I’m not sure I need it on a relatively sedate home harness. Speculative drafts. An invitation arrives, and the house checks the calendar and holds a draft reply before anyone asks. My outbound gate already has a place for it to land. But I don’t know if a low throughput harness even needs this. Maybe not. But I have a version of this that J call proactive check-in.

Pointed the synth boys to this and the tool was tested before I reached work.. haha “for the geometric part of the job. It is scriptable in a handful of lines, runs in seconds, found the side-core undercut unaided, and caught the sliver our own grading had missed. That last one is enough to put it in the benchmark pass permanently” Link to the test process and results in the reply.

Part 2 of the AI CAD Bench. This round covered the other two part families: formed sheet metal and an aluminium extrusion. And it got boring: two kernels, two agents that never saw each other's work, and the exact same 94,544.257 mm³. Same setup as part 1. A fresh Opus 5.5 agent rebuilds a part from an engineering drawing in FreeCAD 1.1.4 and in SolidCraft 0.27.1, from one shared reading of the sheet, with the same rules, graded against the original geometry. Sheet-metal bracket (NIST PMI test case 3, graded against NIST's own model): FreeCAD: 99.65% overlap, volume −0.04% SolidCraft: 99.67% overlap, volume −0.07% Both were missing the same 658 mm³, which is the "NIST" engraving that the rules said to leave out. SolidCraft needed 3 script runs to FreeCAD's 1. Two of those were the agent's own bugs. The third is the fun one: the agent made each slot as a box with a cylinder touching each end, and SolidCraft won't cut where curved faces only graze the body. FreeCAD took it without complaint. A human would never build a slot that way. The agent switched to SolidCraft's slot command, and it worked first time. One rerun. SolidCraft's strictness also helped. It refused a malformed sketch loop with a named error, right at the sketch. FreeCAD might have quietly turned the same wire into a solid that looked right and measured wrong. Heat sink, with nine unevenly spaced fins and twelve holes: identical to three decimal places. Both had 53 faces and fins 1.75 / 1.51 mm wide, 34 s vs 32 s, zero refusals. Both agents also hit on the same workaround on their own: draw one fin, place nine copies by coordinate, fuse once, cut once. All four parts: three ties and one clear FreeCAD win, on the drafted, lofted moulded bracket (99.5% vs 97.1%). So on most parts that get modelled from a drawing, the kernel doesn't decide it. Where it does is tangent booleans and tight fillet chains on drafted faces. I'm curious which of those agents learn to route around first. Caveats: single runs, one model, no sheet-metal workbench on either side, and the heat sink was graded against the reading rather than a model.