backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
Hand-write CLAUDE.md/AGENTS.md + skills, then delete as much as you add — Theo says the audit was 100% worth it — Direct validation of light-prompt / anti-bloat doctrine for your Opus-5 conductor; bloated memory files actively hurt (replies cite Opus inventing “terrorist rules”). — x.com/theo/status/2082009220631953782
EMERGING: “unhobble the model / delete everything / let the model express itself” as the harness design brief — swyx (citing @bcherny + @trq212) reframes agent harnesses as subtractive surfaces, not prompt piles — aligns with strip-skills doctrine and outranks feature news. — x.com/swyx/status/2082104987875004713
Opus 5: thorough + “correct-looking,” but over-literal and turns every comment into a P0 thousand-line fix — Keep tight leash on conductor/review seats; scope boxes and “do not expand” gates matter more than more CLAUDE.md. — x.com/theo/status/2081880182936502474 · x.com/theo/status/2081880547559907675
Security boundary: container isolation is not enough — agents can kernel-panic the host; Firecracker microVMs are the bar — Any unattended loop / sandboxed executor you run should not trust Docker-only isolation if agents get shell. — x.com/rauchg/status/2081842439304995169
Grok 4.5 wins cybersecurity price/perf (DeepsecBench); Sol still frontier, Opus 5 behind Sol; Fable refusal rate “too high” — Route security-adjacent review to GK when budget binds; keep Sol/CX for hard vuln hunts; Fable stays reserve not daily. — x.com/rauchg/status/2081852481517318560 · x.com/rauchg/status/2081855367550751043
EMERGING: persistent agent “brain” outside the sandbox — cloud gateway + nodes everywhere (“claws”) — steipete names the pattern for always-on orchestrators that hop local Mac / VPS / cloud without dying with the shell. — x.com/steipete/status/2081888746497364389
Anthropic (Claude Platform eng) still runs ~6–8 person teams — not “two humans + a bunch of agents” — Calibrate multi-agent ambition: agents amplify squads, they don’t replace fleet design / human review seats. — x.com/GergelyOrosz/status/2082133116026569019
Don’t spend $1M tokens rebuilding buyable software (Jira-class) just because agents can — Prefer agent leverage on your unique stack; buy commodity tools instead of infinite self-build loops. — x.com/kunchenguid/status/2082149546352902428
Bonus signal (lower action): swyx notes Claude Code’s accidental “open source” changed ~nothing on roadmaps — harness source isn’t the moat. — x.com/swyx/status/2081890955070980416
Highest-leverage action today: Spend one focused session auditing CLAUDE.md / AGENTS.md / skills — rewrite the load-bearing few by hand, delete the rest (Theo + unhobble).