Cosmic Rundown: Terry Tao on Coding Agents, Odin Language, and Distributed LLMs

Terry Tao on Coding Agents

Tao's post covers his experience using modern coding agents to build real software, not as a novelty but as a serious workflow. His conclusions: agents are genuinely useful for a specific class of task, and the hard part is giving them structured, reliable context to work with. The post is worth reading in full.

Ant: A New JavaScript Runtime

A new JS runtime called Ant launched today with 305 points and 135 comments on HN. The runtime space keeps moving (Bun, Deno, now Ant), which tells you something about how seriously developers are evaluating their toolchain foundations.

Mesh LLM: Distributed AI on Iroh

Mesh LLM (301 pts) is building distributed AI compute on top of Iroh, a peer-to-peer networking library. Local and distributed model inference is a real and growing category. The implication for builders: your stack should be able to route to whatever compute makes sense, not locked to one provider.

An Agent in 100 Lines of Lisp

A clean implementation of an agent loop in 100 lines of Lisp (196 pts) making the rounds again. The point: the agent loop itself is simple. The hard part is what the agent acts on, the structured, governed, reliable backend that determines whether agent output is useful.

Mindwalk: Replay Coding-Agent Sessions

Mindwalk (119 pts) lets you replay coding-agent sessions on a 3D map of your codebase. Agent observability and auditability keep becoming more important as agents do more real work.