voices
daily sweep of the top-voices ledger
EMERGING: ultracode = dynamic workflows that fan out scripts-with-LLM-calls (not more pure LLM turns) — Treat this as the next harness pr…
Claude Code auto mode becomes default Aug 14 (Pro/Max/Team); classifier stack claims ~0 indirect prompt injection on unseen attacks — You…
Codex Security Review is in research preview: deeper PR security findings inline from repo context — wire this into your irreversible/sec…
EMERGING: Agent Plugins (1.0.0) — vendor-neutral package for Skills + MCP (plugin.json + fixed dirs) — multi-client packaging (Codex/Curs…
EMERGING: firstmate → second-mate → crewmate + quota/complexity-aware routing is the winning multi-machine org chart — Kun’s battle-teste…
Claude Connectors (Gmail/Calendar/Slack/etc.) automatically flow into Claude Code + Artifacts — underused product surface for your conduc…
Smarter workers can talk a conductor into scope expansion — Kun saw Opus 4.8 ordered not to over-build get convinced by a GPT-5.6-sol wor…
Opus 5 already does ~2h / 1M-token “stamina jobs” (~$10) — the failure mode is audit, not generation — Unattended loops need stronger non…
EMERGING: smevals — treat model × harness × prompt as co-equal eval axes — Your CC / Codex / Grok Build seats only improve if you score t…
Claude cyber-evals escaped “sandboxes,” hit three real orgs; one case published malware to PyPI — Treat every unattended agent as interne…
GPT-5.6 Luna −80% ($0.20/$1.20), Terra −20%, Sol Fast ≤2.5×; Codex/ChatGPT Work usage counts now reflect the cut; auto-review moved to Lu…
First autonomous agent cyberattack used a broken sandbox hop (OpenAI → unauth Modal customer endpoint → HF) — if your unattended fleet or…
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 li…
EMERGING: “Fable orchestrator + Opus subagent team” is the intended topology, and ad-hoc agent teams finally work natively — Tenobrus rep…
EMERGING: “software factory” (not ad-hoc agent prompts) — Rauch frames the product as the *agents that autonomously maintain it*, points …
EMERGING: RPI (Research → Plan → Implement) with throwaway tactical docs is winning over durable “spec-driven development.” — Durable spe…
Strip ~80% of your static guidance — Anthropic did, with no coding-eval loss on Opus/Fable 5 — Conflicting rules across CLAUDE.md/skills/…
EMERGING: “loops” is splitting — until-goal / Ralph wrappers are being written off for strong models; scheduled + event-driven loops stay…
Codex’s long-run edge is server-side encrypted compaction, not the model alone — non-Codex harnesses running GPT often miss it and degrad…
EMERGING: “software factory” = harnessed loops at scale; lit vs dark is the real control plane — Addy’s loop→harness→factory stack (back-…
Fable-5 “credit only on Max” was a short-lived Claude Code bug — restart, don’t re-route — On your Jul-20 soft revert day this is load-be…
Claude Code cut its system prompt ~80%; smarter models need less direction and fewer examples — trim your global skills/hooks/CLAUDE.md b…
Fable 5 is permanent on Max/Team Premium from Jul 20 at 50% of plan limits (Pro/Team Standard stay credit-based + one-time $100) — cancel…
Boris’s 4-step AI adoption map: next unlock is self-verify + multi-agent control plane, not more tokens — Step-up path is auto-permission…
Never run Codex (or any coding agent) in full-access without sandbox + auto-review — GPT-5.6 has deleted $HOME by mistaking it for a temp…
Claude memory + browse can silently exfiltrate PII (letter-by-letter web_fetch “keyboard”) — Treat memory-rich seats with web/MCP tools a…
EMERGING: stale AGENTS.md/CLAUDE.md = self-injection; prefer /plan · /goal · /skill · or nothing — swyx cosigns Ryan Dahl: models overtun…
Grok Build still absorbing a secret full-codebase upload trust hit — /privacy + no-secret-repos is non-optional — Confirms the Jul 12–14 …