voices

Voices digest

Aug 1, 2026

backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 24h


EMERGING: smevals — treat model × harness × prompt as co-equal eval axes — Your CC / Codex / Grok Build seats only improve if you score them on the same small task suite, not vibes; Simon’s vocabulary puts the harness (not just the model) on the Pareto front. — x.com/simonw/status/2083310510729216039

  1. Production pattern: /loop fuzzer → agent fixer (Jarred / Bun) — Closed-loop unattended QA: a Claude Code /loop drives a syscall-diff fuzzer and pipes failures to robobun to patch — the template for your verification gates. — x.com/jarredsumner/status/2083542090953470212

  2. Opus 5 can torch a 5h Claude window in ~10 min via spawn storms — Theo hit the limit when the agent spawned a checker every 2 seconds; daily spawn caps alone don’t stop high-frequency subagent thrash. — x.com/theo/status/2083296221159121156

  3. GPT-5.5 / Codex: mid-run steering beats queueing — steipete: with 5.5 you can throw work at a live agent and it processes it without the old queue discipline; OpenClaw’s queue/steer path is reworked for a release next week. — x.com/steipete/status/2083369880599015713

  4. Stateless MCP is the deploy story — Simon rekindled MCP interest around the new stateless spec (mcp-explorer, datasette-mcp); session affinity was the blocker for cheap Workers/Lambda-style MCP hosting. — x.com/simonw/status/2083330693313220615

  5. Fleet economics: 3×$200 Claude + 1×$200 Codex, meter with ccusage-fleet — Theo’s ~$50k API-equivalent July spend (mostly Fable + gpt-5.6-sol; ~$20k “work that mattered”) and multi-machine ccusage aggregation match your C1/C2/CX topology. — x.com/theo/status/2083349476606120116

  6. Kun Chen open-sourced full Pi agent config (dotfiles + walkthrough) — Concrete peer-harness settings from the no-mistakes author; steal structure for skills/hooks/agent defaults, not brand loyalty. — x.com/kunchenguid/status/2083611681868107857

  7. Agent-era gates: more agent-maintained E2E, less stacked-PR theater — Gergely: agents make E2E cheaper to write/fix so volume rises (determinism still the prize); stacked diffs lose leverage when models author most of the diff. — x.com/GergelyOrosz/status/2083469238417723808

Action today: Add a hard spawn-rate / checker-frequency cap on Opus 5 subagents (not only daily count) so a 2s-loop can’t burn a 5h window.