backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
EMERGING: firstmate → second-mate → crewmate + quota/complexity-aware routing is the winning multi-machine org chart — Kun’s battle-tested stack (Grok 4.5 orchestrator, remote second mates over Tailscale/SSH, crewmate model picks by capability + ambiguity + remaining quota, no-mistakes on GPT-5.6 sol) is a direct upgrade blueprint for your conductor/executor fleet. — x.com/kunchenguid/status/2084825003099607189
GPT Luna is “basically free” after an ~80% price cut; gdb agrees on price-performance — shift bulk summarization, status lines, tool-call digests, and other non-coding agent glue off Max seats onto Luna so Claude/Codex quota stays for real code. — x.com/theo/status/2084748639470272972 · x.com/gdb/status/2084770040281211043
Give coding agents eyes+hands on real hardware (video KVM), not only VMs — steipete’s Codex-on-remote-KVM e2e for OpenClaw/iMessage is the same verification gap your cua / real-device loops exist for; second account needed for true e2e. — x.com/steipete/status/2084988316324397312
GPT models “I’m going to do X” then idle is a known harness failure; codex-oss has the nudge pattern — steipete/mitsuhiko: Pi hit this; system-prompt + harness nudges (Claude too) matter; rebuild on app-server fixed OAI for them — audit your stop/continue gates so executors don’t announce without acting. — x.com/steipete/status/2084990946937340164 · x.com/mitsuhiko/status/2084995000673218898
Anthropic Agents SDK still TOS-legal for wrappers; planned clamp “delayed indefinitely” — Theo (T3 Code) + Kun (Pi for non-Claude, Claude only under Anthropic TOS) means keep multi-harness but don’t assume Claude CLI will stay open as a free orchestrator host for other models. — x.com/theo/status/2084892066321244475 · x.com/kunchenguid/status/2085032264459813023
Claude Managed Agents = Agent-SDK benefits with Anthropic-owned sandboxing — trq212 points product surface for long-horizon work when you want less DIY harness/sandbox; relevant alternative to home-grown isolation. — x.com/trq212/status/2084782162033910216
Codex activity view improved but still not daily-driver UI — Theo still prefers T3 Code sidebar; don’t retool your control plane around Codex activity yet. — x.com/theo/status/2084801515689316700
simonw LLM CLI big release: reasoning traces, OpenAI Responses, server-side tools, better logging — useful if you need a thin multi-provider adapter layer with debug-grade traces outside Claude Code/Codex. — x.com/simonw/status/2084792341572001871
EMERGING (antenna): ledger voices (Ng, Orosz, swyx, Osmani) amplified Jeff Dean et al.’s Discovery Loop — “loop” language moving from coding harnesses up to lab-scale automated science/ML/engineering discovery. — x.com/AndrewYNg/status/2085056542341271840
Today’s action: Encode Kun-style routing rules (capability × ambiguity × live quota) into your conductor dispatch, and move status/summary/glue work onto Luna.