backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
EMERGING: “Fable orchestrator + Opus subagent team” is the intended topology, and ad-hoc agent teams finally work natively — Tenobrus reports Fable/Opus/Sol can split 10+ agents across worktrees with little extra scaffolding, but partial specs now waste architecture-scale work (treat DoR as mandatory before fan-out). — x.com/tenobrus/status/2081517854302851444
Opus 5 is a “natural born subagent,” not the default human-facing conductor — same thread family: technical/subagent-strong, rigid at high-level ideation; matches seating Fable/Opus-5 conductor + Opus executors rather than Opus-as-sole interlocutor. — x.com/tenobrus/status/2081462333159952842
Boris/YC: delete ~80% of system prompts, “unhobble,” run thousands of agents — fresh Startup School talk is the primary-source pass on strip-prompt / harder problems / multi-agent at scale; treat Fable-era skill bloat as active drag on Opus 5. — x.com/ycombinator/status/2081787356420718704
EMERGING: harness absorption — “the model is the harness” — Knoop: Claude Code/Work harness behavior is being trained into Opus 5 raw; CC users under-see the jump because the harness already did the job → stop re-implementing ReAct/subagent scaffolding the model already internalized. — x.com/mikeknoop/status/2081792754309587333
Opus inherits the GPT-5.6-sol “scope creep” failure mode; Fable still wins at slop-trim — Theo: Opus piles on +100 LOC fixes / +200 LOC tests; keep Fable as reserve trim/taste seat after Opus execution. — x.com/theo/status/2081551375985709139
Cross-org agent→agent loops are live ops, not demos — steipete’s agent reported a Bun bug overnight; jarredsumner’s robobun agent fixed it the same night — pattern for unattended issue/PR interop across fleets. — x.com/steipete/status/2081767828278170002
Nanocodex + harness-latency doctrine (Codex seat) — gakonst: headless, extensible Codex library (no cruft); separately: harness must be network-latency-bound (2s startup is unacceptable); skills called more distracting than helpful for this shape of agent. — x.com/gakonst/status/2081753611693109378 · x.com/gakonst/status/2081595520284508311 · x.com/gakonst/status/2081799031060304319
Subscription packaging split: Claude meters speed, Codex meters context — Theo: both offer “fast mode” + ~1M ctx; Claude sub includes 1M, charges for fast; Codex includes fast, charges for higher context — route long-horizon/context work to Claude Max, latency-sensitive bursts to Codex. — x.com/theo/status/2081549009731420566
Highest-leverage action today: before any multi-agent fan-out, write a hard Definition of Ready (goal, write-set leases, done_when, out-of-scope), then run a delete-pass on CLAUDE.md/skills keeping only repo-unknowable gotchas.