Velocity News

xAI just shipped Grok Build — a $300/month terminal coding agent that reads Claude Code's own config files and runs Grok 4.3. Today's video is the first hands-on, including the day-one bug nobody at xAI is talking about. Five repos that map the agentic-CLI landscape it just walked into: a DeepSeek terminal agent that gained 21k stars in a week, a semantic-search MCP server every coding agent can share, and the cross-session memory layer that fixes the "new session, lost context" problem. A benchmark rail on where Grok 4.3 actually lands, and a config-format note worth reading before you let any agent touch your repo.

Monday. Today's video tackles the problem every one of us hit last week — AI coding tools choking on real codebases and burning tokens to do it. Five repos, three of which are about making LLM pipelines actually reliable in production rather than impressive in a demo. The bench table is refreshed after a quiet weekend. And the caching change we flagged twice last week goes live today — see Heads up.

Sunday edition. Google split its 8th-gen TPU into two chips — one for training, one for inference — and the architectural rationale is a 10-minute crash course in why agentic workloads are bending hardware. Five repos that earned today's slots, including Salvatore Sanfilippo's Metal-native DeepSeek 4 Flash engine, Anthropic's just-launched financial-services agent skills, and SGLang quietly crossing 27k stars. A paper on TPU-aware compiler optimizations to pair with the video, and a benchmark observation worth noting before Monday.

Sunday edition. Nous Research's Hermes Agent — coincidentally sharing a name with our own meta-layer roadmap item — just shipped a desktop app, and the video walking through it is worth nine minutes of your evening. Five repos that genuinely earn their place, including one that's quietly become the production backbone for half the AI startups you know. The bench table is refreshed; one frontier model moved overnight.

It's Saturday, which means today's edition tilts toward thinking rather than shipping. A 30-minute talk from AI Engineer that just premiered, on why the engineers building well with AI agents aren't the ones using fancier tools — they're the ones falling back on fundamentals nobody talks about anymore. Five fresh repos worth your weekend. The bench table refreshed with this morning's leaderboard movement. Read it with the second cup of coffee.

It's Friday. Five repos worth your weekend, a fresh comparison of where every frontier model actually stands today, a paper that made us reconsider how we structure agent memory, and a video on whether Kimi K2.6 + OpenCode at $10/mo is the Claude Code alternative our budget calls have been asking about. Reading time: under five minutes.