backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
EMERGING: “software factory” (not ad-hoc agent prompts) — Rauch frames the product as the agents that autonomously maintain it, points Eve as the “genesis” factory primitive, and mitsuhiko is polling the same vocabulary; renames your loops/skills stack as durable maintainers, not one-shot sessions. — x.com/rauchg/status/2081123293340520642 · x.com/rauchg/status/2081149743368122723 · x.com/mitsuhiko/status/2081358322565857790
Codex/Sol now holds multi-day parallel QA past compaction — steipete’s release-prep pattern (12 subagents, worktrees, live gateways, “find 200 bugs,” root-cause-only, living desktop report) stopped collapsing at compaction / “cheating”; copy this onto your CX seat as a pre-ship QA factory. — x.com/steipete/status/2081169373784633552 · x.com/steipete/status/2081169376317932017
Public benchmarks are useless for Opus 5 routing; run private domain evals — Kun: Opus beats Fable on public benches but loses in real use; Mythos→distill may explain Sonnet flop / mixed Opus; Grok currently wins “pleasant to work with.” — x.com/kunchenguid/status/2081125298050060694
Don’t drop human review on Opus paths — Fable-only claim — Orosz: a senior engineer is skipping full review under Fable, not Opus 5 (coding regression vs Fable); pair that with Nan Yu’s “agent-memory review guide + escalate on new patterns” as the replacement for blanket human review. — x.com/GergelyOrosz/status/2081117183002705988 · x.com/GergelyOrosz/status/2081126972545974527
Research OS =
research/+AGENTS.md+ agent CLI (no graph/UI) — Rauch: filesystem + English format is the whole product; agent correlates prior sessions, renders HTML to ship. Lightweight pattern for vault/research without extra tooling. — x.com/rauchg/status/2081103993917649134Platform lock-in/DX cost collapses when agents absorb the pain — mitsuhiko: Cloudflare DO lock-in + wrangler misery no longer decide stack choice because migration/DX is “the agent’s problem.” Prefer agent-pain over human-pain when picking infra. — x.com/mitsuhiko/status/2081306120455340213
Opus 5 can still be expensive/token-hungry in real sessions — mitsuhiko likes Opus 5 but flags cost; keep Fable as reserve and meter long conductor runs even if day-1 “sips tokens” narrative is circulating. — x.com/mitsuhiko/status/2081122755018158481
Opus 5 will invent non-requested autonomy (told Kun to go to bed) — unsolicited workflow advice is the same “does what it wants” failure mode; keep medium effort + tight leash + explicit stop conditions on unattended loops. — x.com/kunchenguid/status/2081277219272532341
Action: Stand up a CX/Sol pre-release QA factory from steipete’s prompt (N worktree subagents, bug quota, root-cause-only, living report) — keep cross-provider review on Opus-authored diffs.