Frontier scan

backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 84h

Ranked (max 8) — by how much each would change your ops

1. Harness×model pairing as first-class design (not “which model” alone)
   What's emerging: Same frontier model is treated as different tools per harness; power users route brains through the better shell (often Claude Code as conductor, Codex/Grok as peer seats or plugins).
   Evidence: Independent practitioners report GPT-5.6 performing “meaningfully better in Claude Code than in Codex,” and others consolidating active harnesses while wiring Codex/Grok as managed subagents inside Claude Code rather than multitool hopping.
   Replaces: Treating harnesses as interchangeable UIs; equal multi-app daily driver stacks.
   URLs: x.com/DevAutomata/status/2077022846337859951 · x.com/TSpencer260/status/2077172564468625669 · x.com/IanAndersonLOL/status/2076540263007457477

2. Loop engineering with explicit stop conditions (autonomy ladder)
   What's emerging: Work is designed as goal-/time-/event-triggered loops with budgets and evaluators; human jobs (start + stop) are what get automated, not longer prompts.
   Evidence: Multi-source framing this week: stop-condition taxonomy (exploratory → measurable → recurring → standing), practitioner architecture posts (automations + worktrees + skills + connectors + non-author subagents), plus mainstream engineering podcast airtime.
   Replaces: Turn-based “prompt engineering” / chat-shaped agent sessions.
   URLs: x.com/Degen_calls_sol/status/2077298166211256321 · x.com/0xbelorix/status/2076893702941810927 · x.com/GergelyOrosz/status/2077434907274428914

3. Context hygiene as production discipline (“dumb zone,” trajectory poison, intentional compaction)
   What's emerging: Session quality is managed by what stays out of context; operators kill sessions on sycophancy cues and budget state files like memory management.
   Evidence: High-signal Orosz↔dexhorthy episode codifies “dumb zone,” intentional compaction, and “You’re completely right!” = restart; separate multi-session operators report state-file bloat → quality drop misread as model drift.
   Replaces: “Fill the window” habits, endless sessions, treating compaction as free.
   URLs: x.com/GergelyOrosz/status/2077434907274428914 · x.com/dexhorthy/status/2077438225149505599 · x.com/johniosifov/status/2076816228472295530

4. Shell/tool-output compression as a standard layer (RTK-class + MCP gateways)
   What's emerging: A new infrastructure category between agent and tools: compress/filter shell (and aggregate MCP tools) before tokens hit the model.
   Evidence: Real open-source tooling with multi-month install path (rtk init -g hooks for Claude Code/Codex/etc.) and ongoing practitioner demos; honest caveats (Bash-hook only; Read/Grep/Glob bypass) still circulate—adoption signal stronger than pure launch spam.
   Replaces / threatens: Raw shell dumps and naive multi-MCP tool-surface dumps as default context feeds.
   URLs: github.com/rtk-ai/rtk · x.com/RetroChainer/status/2077459489570046346 · x.com/fluixoo/status/2077309853731561801

5. Advisor/executor split standardized as plugins (not custom glue only)
   What's emerging: Expensive model plans / judges; cheaper model implements—now shipping as marketplace plugins on both Claude and Codex sides.
   Evidence: Codex-orchestration advisor/executor install flows are being set up by independent users in minutes; OpenAI’s official Claude-side Codex plugin continues to push review/rescue as peer roles.
   Replaces: Single-model do-everything sessions; ad-hoc “call the other CLI” scripts without role split.
   URLs: x.com/Av1dlive/status/2077016063578161578 · x.com/RoundtableSpace/status/2077185123137560897 · x.com/kirillk_web3/status/2076617885506740408

6. “Less is more” harness + skill stack (plugin pollution backlash)
   What's emerging: Power users prune: few harnesses, thin skill sets, prefer harness-native commands over mega plugin packs that bloat context.
   Evidence: Practitioners explicitly cite cognitive overload and settle on Claude Code + one secondary shell with selective plugins; others call superpowers-style packs mostly obsolete now that /code-review, /simplify, /verify, /doctor live in-harness.
   Replaces: “Install 154 agents / 1000 skills” maximalism (still loud on X—treat as hype unless you curated).
   URLs: x.com/TSpencer260/status/2077172564468625669 · x.com/blocktivist/status/2076599095624339559 · x.com/morpphhhaw/status/2077392175705317796 (hype counter-example)

7. Portable agent definitions as markdown+YAML (skills → custom agents everywhere)
   What's emerging: Agents/subagents defined as files with frontmatter, launchable via CLI flags; same pattern spreading beyond Claude Code (e.g. Antigravity).
   Evidence: DeepMind-adjacent shipping note: custom agents/subagents as markdown+YAML + --agent flag; ecosystem posts treat skills as the cross-agent packaging unit (Solana skills for Codex/Claude/Cursor).
   Replaces: Proprietary per-IDE agent configs; one-off system prompts that don’t travel.
   URLs: x.com/JackWoth98/status/2077459614933692780 · x.com/GuiBibeau/status/2077377906939195443 · github.com/solana-foundation/solana-dev-skill

8. Skills as the ops interface to specialized infra (beyond coding assistants)
   What's emerging: “Agent skills” becoming how coding agents drive domain stacks (ML training, domain SDKs)—not just repo conventions.
   Evidence: NVIDIA reports Codex + TAO agent skills autonomously running baseline → LoRA → multi-node AutoML to large accuracy jumps; product/eng posts treat skills as the specialization layer over general agents.
   Replaces: Manual MLOps notebooks / human-driven experiment scripts for some pipelines; one-size coding agents without domain packs.
   URLs: x.com/kimmonismus/status/2077400362995388729 · firecrawl.dev/blog/best-ai-coding-agents · danielkeller.com/tech/agentic-stack-skills-harness/

Terminology shifts (this window)

New voices (not on your ledger)

Skipped as mostly hype amplifiers this window: mega “154 agents / 1000 skills” list accounts, Codex Micro hardware stunts (product PR, not practice).

Tool-category shifts

Ledger proposals

Add

• @dexhorthy — highest-density ops language for the stack you already run (loops/context/harness).
• @_philschmid — empirical harness engineering; non-Anthropic lab view.
• @thdxr — OpenCode peer harness; catches defaults your Claude-centric ledger misses.
• @simpsoka + @romainhuet (pick one seat if slot-limited: simpsoka) — Codex product reality vs rumor.
• @TSpencer260 — multi-pool harness consolidation practitioner signal (not vendor).

Drop / demote

• No forced drops from your named 20 this window on absence of signal—but watch for staleness if any are pure product announce accounts vs weekly operational posts.
• Do not add viral “awesome-*-skills” amplifiers unless they ship original runbooks with failure modes.

No change required on: karpathy, bcherny, steipete, kunchenguid, GergelyOrosz (still load-bearing this week—Orosz platformed the dex episode).

Skepticism tag: Codex Micro hardware, 154-agent marketplace posts, and “71k stars / 97% tokens” claim chains look coordinated/viral; RTK and advisor–executor plugins show independent installs + honest limits → keep those. Loop/harness/context language is no longer niche branding—it’s how serious operators describe the job.