01 — the system
what makes it agentic
Ξ
Policy, not script
A chatbot answers; an agent acts. What separates them is a policy — a way of choosing actions under uncertainty, scored by consequences, not applause.
Ω
Tools with boundaries
Capabilities arrive as instruments with edges: explicit inputs, honest outputs, and permissions you can enumerate. Power you can audit, not magic you trust.
Ψ
Loops that keep score
Every run gets evaluated; every evaluation feeds the next run. Iteration is the only credential a system earns — watch the lattice think and you are watching one.
02 — the harness
where agents live
i.Portable skills — capabilities as files, not folkloreworkshop
ii.Multi-agent orchestration — specialists, not a crowd of oneworkshop
iii.Pure-MLX models — small, sharp, entirely localapple silicon
iv.MIT, in public — the harness is open because trust should be inspectableappautomaton
03 — the interface
where humans meet agents
α.Not a chat window — a surface that shows the workagentic ui
β.Plans you can inspect, tool calls you can vetoagentic ui
γ.Trust through legibility — elegance is a readable systemagentic ux