1/
I'm building motors & sensors in public — an open-source, local-first harness that makes your coding agent get better the more you use it. Not a new agent; a layer on the one you already run (Claude Code / Codex / Gemini). How it works 🧵
Today I'm publishing a new essay, Policy on the AI Exponential. AI is progressing extremely fast—much faster than the policy process was built to handle. The essay lays out where I think the technology is now, and the action needed to close the gap: https://t.co/Lh6PWae178
Is SF that good? It most probably is.
I feel so left out and given the 100K entry lot - it's so difficult to even start the plan for SF dream.
Maybe maybe not..
harness diary · day 2 — motors & sensors
yesterday: 4 hosts traced. today: made it installable.
→ npm i -g motorsandsensors — zero deps, 33 kB
→ mns init scaffolds your agent's faculty home (.mns/) — works like git init
→ repo went professional: 4 living docs, one build journal, one test home
→ publishing is now token-free (npm OIDC trusted publishing) + provenance-attested and the pipeline earned its keep in hour one: first CI run caught a Linux bug my mac was hiding. 0.1.1 shipped the same hour.
lesson: the registry + CI find what your laptop hides.
harness diary · day 1 — motors & sensors
this morning: design docs, zero code. tonight: pushed to github: https://t.co/VtwYzwGkQb
→ host-agnostic trace capture (OpenTelemetry): any coding-agent session
→ a trace you own
→ mns, a git-native CLI — enable once, capture goes invisible
→ 4 hosts, all verified on REAL sessions: Claude Code · Gemini · Codex · OpenCode
the bet: wrap the agent you already run, never drive it. one tiny adapter per host, zero changes to the core — proved it 4×.
biggest lesson (full thread Thursday): the vendor docs lied twice. build adapters from real wire
data, not docs.
harness diary · day 1 — motors & sensors
this morning: design docs, zero code. tonight: pushed to github: https://t.co/VtwYzwGkQb
→ host-agnostic trace capture (OpenTelemetry): any coding-agent session
→ a trace you own
→ mns, a git-native CLI — enable once, capture goes invisible
→ 4 hosts, all verified on REAL sessions: Claude Code · Gemini · Codex · OpenCode
the bet: wrap the agent you already run, never drive it. one tiny adapter per host, zero changes to the core — proved it 4×.
biggest lesson (full thread Thursday): the vendor docs lied twice. build adapters from real wire
data, not docs.
1/
I'm building motors & sensors in public — an open-source, local-first harness that makes your coding agent get better the more you use it. Not a new agent; a layer on the one you already run (Claude Code / Codex / Gemini). How it works 🧵
4/
And it's local-first — the whole thing runs on your machine, nothing leaves by default. Open source from commit zero. I'll post the wins, the dead-ends, and the learnings here as I build. Follow along 👇
1/
I'm building motors & sensors in public — an open-source, local-first harness that makes your coding agent get better the more you use it. Not a new agent; a layer on the one you already run (Claude Code / Codex / Gemini). How it works 🧵
3/
The twist: these faculties graduate. It watches your real runs (the sensors), scores how each one performed, and proposes upgrades — a sharper tool, a promoted memory, a tighter guardrail. You approve. It pins a new generation. Rollback = flip a pointer.
@im_roy_lee seriously?
no 'cheat-coder' or 'vibe-coder' or even 'vibe-operator' (non-engineer) role.
I'm building an open-source tool for same - https://t.co/WTNgxEnkPn
hire me on trial and fire me if you don't see value - [email protected] or twitter dm.
Re-inventing education program applications are open now!
Want to know how turn your ideas for education into an impactful solution? Want to re-think the status quo in education? Want to learn from the best in this sector? Maybe be a part of a community of like minded individuals?
If all your answers were a "yes" then this program is for you! APPLY NOW 💥