backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
Claude Connectors (Gmail/Calendar/Slack/etc.) automatically flow into Claude Code + Artifacts — underused product surface for your conductor/executor seat: one connect path wires personal/ops context into terminal agents, not just chat. — x.com/trq212/status/2084387303959740449
Claude still refuses
AGENTS.md(and peer tools still fight over skill roots like~/.agents/skills) — keeps your dual-doc / symlink tax load-bearing; don’t assume one agents-file is enough for cross-seat (Claude vs Codex) runs. — x.com/mitsuhiko/status/2084629231963619660OpenAI staff: Codex is “a good harness” now, but will look primitive in 2–3 months; next models need more than a laptop — treat CX as a peer harness to re-eval on a short half-life, not a fixed architecture bet; expect cloud/remote compute as the next harness layer. — x.com/reach_vb/status/2084487109373530545
EMERGING: “skill bankruptcy” — wipe local skills/plugins and rebuild only what still earns its keep as models + harnesses improve — matches your GC doctrine; Addy keeps only skills that still make a measurable diff. — x.com/addyosmani/status/2084505157442703554
EMERGING: ephemeral experiment driver — one unit of work = one unit of compute (prompt-spawned VMs; “convention as product”) — Huntley day-2 full-time on
exe.dev/ shelly for sandboxed experiments; stronger isolation pattern than shared-machine agent runs. — x.com/GeoffreyHuntley/status/2084686743606075818Price value created against true serving cost (+ margin); $1M/yr tokens on a side project is not the future — reinforces hard pool caps and “cheapest sub-agent is the one you never spawned” for unattended loops. — x.com/mitsuhiko/status/2084563924435845350
Taste can be borrowed from agents; judgment cannot — judgment is liability (name on the call), not a capability to transfer — keeps human gates on irreversible/security merges; agents rank options, humans own ship/no-ship. — x.com/addyosmani/status/2084354578196443351
GPT models often end with “I’m going to do X” and then do nothing — intention text ≠ tool execution — for CX/Codex-adjacent seats: require tool-call / patch / test evidence in gates; never treat plan prose as completion. — x.com/mitsuhiko/status/2084673337733640301
Also noted (lower leverage): remote agent-browser + vercel i kernel|browserbase path for headless agent browsing (https://x.com/rauchg/status/2084345210470097127); OpenAI DevEx team actively collecting what to build (https://x.com/reach_vb/status/2084671716425675206); laptop-class Qwen 3.8 on deck (https://x.com/simonw/status/2084667167212245170). Quiet: bcherny, indydevdan, jarredsumner, _catwu, AndrewYNg, delba_oliveira.
Action: Turn on only the Claude Connectors you’d trust in Code (or document an explicit deny), then run a skill-bankruptcy pass — delete zero-diff skills so the connector context budget isn’t wasted on stale harness prose.