self.md radar — 2026-10-03
three ways the machine stays after you close the tab
Today’s useful detail sits below the chat window: a local service on port 5476, kernels tuned before a model runs, and a motion dataset that still asks you to buy the animal pack. The demos are familiar. The plumbing is where the bill starts.
1. Kiro Crew puts schedules and heartbeats inside a persistent dev workspace
Kiro Crew is an Apache-2.0 workspace for development work that is meant to survive the end of a session. Its README puts desktop, web and CLI access beside unattended multi-step tasks, recurring jobs and heartbeats that watch systems until something needs attention.
The quick installer ends by opening localhost:5476. The default agent still depends on a separately installed and signed-in Kiro CLI. That small line is more honest than the word “persistent”: someone still owns the machine, the service and the login.
reading: Kiro Crew repository
2. Magnitude tunes its kernels on the device, then makes a speed claim
Magnitude is a new open-source inference engine for agent use. Instead of shipping only broad hardware targets, it says it compiles and tunes kernels on the actual Apple Silicon, NVIDIA, AMD or CPU machine before a model runs.
Its README reports 92% faster decode on Metal and 19% on CUDA against llama.cpp, plus 27% less memory per agent. Those are the project’s own numbers, not an independent bake-off. Still, the product choice is clear: local performance is being sold as a per-device calibration job, not just a model download.
reading: Magnitude repository
3. UniMate’s one model still arrives with a list of things you must source
UniMate released training and inference code on September 6, then preview checkpoints on September 27. The project’s UniML3D dataset has 13,006 text-paired motion sequences across people, animals, birds, insects, marine creatures, serpentine forms and rigid objects.
The repository also names the awkward part: its pipeline uses a commercial Truebones animal-motion pack whose license does not allow redistribution. One unified animation model is useful. It does not turn asset rights, preprocessing and strange rigs into somebody else’s problem.
reading: UniMate repository
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- Local LLM runtimes — pick local inference by custody, speed, privacy, cost, and failure mode
- Agent memory systems — what agents should remember, what belongs in logs, and how to avoid memory sludge