backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h
Claude memory + browse can silently exfiltrate PII (letter-by-letter web_fetch “keyboard”) — Treat memory-rich seats with web/MCP tools as an exfil surface (same pattern reaches Drive/inbox/MCP, not just chat memory); keep untrusted URLs and link-following off high-context conductors. — x.com/steipete/status/2077303292225548539 · writeup: ayush.digital/blog/the-memory-heist
EMERGING: “dumb zone” + “trajectory-poisoned” session language (context/harness/loop stack) — Cap live context before quality cliffs; on “you’re completely right / right to push back,” hard-new-session instead of continuing; unread agent code eventually becomes un-debuggable even for Opus-class models. — x.com/GergelyOrosz/status/2077434907274428914
Codex: don’t
/goalthe raw ask — research first, then let the agentset_goalwith a verifiable success check — Cuts search space and token burn on unattended CX runs; if the agent can’t verify the goal, it will thrash. — x.com/reach_vb/status/2077136521178517999GPT-5.6 Sol failure modes (no-mistakes author): jargon fog, over-engineering spirals, mock-heavy “validation” that fails prod — If CX is peer executor, pin Sol with plain-English, minimal-diff, live-env/real-path verification — don’t assume Claude-shaped defaults. — x.com/kunchenguid/status/2077448372999569557
Claude Code:
Ctrl+Bbackgrounds a long-running command and unblocks the turn (result returns when done) — Direct harness win for install/build/test waits so the conductor keeps routing instead of stalling. — x.com/delba_oliveira/status/2077419708324745412Fable vs Sol plan bake-off: same prompt → cross-critique; Fable reliably prefers Sol’s plan — Prefer Sol (or dual-plan) for planning, Fable for critique — not the inverse “Fable always plans” doctrine. — x.com/swyx/status/2077178561354735962
Fable + gpt-5.6 missed an 85% GPU win (
animate-pulse+ blur + grain); agents helped as toolmakers, not root-cause finders — For perf/UI bugs, force instrumentation/console toggles first; don’t expect either model to own the diagnosis alone. — x.com/theo/status/2077317685679985119Always-on autoreview skill (OpenClaw) — burn tokens to calm nerves after agent edits — Aligns with your automated-gate floor; cheap insurance after unattended loops vs hoping self-review holds. — x.com/steipete/status/2077265627379843242
Highest-leverage action today: Pin CX/Sol (plain language · minimal diffs · no mock-for-prod) and switch unattended Codex to research → self-set_goal with a machine-checkable verifier.