backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 84h
independent practitioner adoption + ability to change how you operate.
1. Code-as-environment over skill/doc bloat
What's emerging: Power users are cutting agent-facing prose (long CLAUDE.md, explanatory docs, skill walls) and spending the token budget on self-explanatory code + thin navigation pointers — skills only where the model truly lacks affordance.
Evidence: Matt Pocock’s “delete the docs… make code self-explanatory” hit ~2.5k likes / ~293k views; replies explicitly say they ship without CLAUDE/AGENTS/skills and treat skills as lab context hacks; Justin Schroeder (StandardAgents) states skills often counterproductive on frontier models and “MCP is just silly… a directory of scripts is better.”
URLs: x.com/mattpocockuk/status/2076596585069174925 · x.com/Shashikant86/status/2076954198424338863 · x.com/jpschroeder/status/2076520407486628253
Replaces: “more skills / fatter CLAUDE.md = better agent” as the default upgrade path.
2. Session-reset / asymmetric verification (verifyer never shared implementer context)
What's emerging: Verification is being formalized as a fresh session that only sees intent + diff, because same-thread agents defend their own work.
Evidence: @koder0x describes a daily work loop (plan → fan-out agents → complexity-pruner → reset session → fact-checker vs DevOps story); independent echo: “reviewing in a fresh sub agent without the session context is a smart trick.”
URLs: x.com/koder0x/status/2075825401058726249 · x.com/rohit_jsfreaky/status/2076707987218121098
Replaces: In-session self-review / “structured self-critique” as the quality gate.
3. “Loop engineering” is peaking as a word — practice stays, brand gets skeptical
What's emerging: Serious builders are questioning the label, not the scheduled/goal loops. Gergely reframes useful cases as event-triggered agents (spec landed → implement; bug filed → fix) — “not all that different from webhooks/Zapier.”
Evidence: Gergely’s poll-style post (~730 likes, ~161k views, 200+ replies) is explicitly “not (yet?) buying that it’s a new paradigm”; same window is flooded with near-identical “Anthropic free loop course” farm posts (coordinated hype signature).
URLs: x.com/GergelyOrosz/status/2076591749309374650 · x.com/GergelyOrosz/status/2076604325741076832 · (hype contrast) x.com/BIGMayrr/status/2076958486240354763
Dying term: loop engineering as paradigm claim. Surviving practice: time/goal/event triggers + evaluators. Rising alternate frame: events engineering / “designing data-intensive agents” (schema memory > loop slogans) — x.com/lwastuargo/status/2076901789459435822
4. Plugins as the distribution unit (skills leave the folder)
What's emerging: SKILL.md is no longer enough — plugins (skills + agents + hooks + MCP as installable packages) and npx skills add … are how packs spread; “grill-me” is organic cross-language adoption.
Evidence: Pocock submitted a Claude plugin for his skill set; third parties install via npx skills@latest add mattpocock/skills and report full skill chains through implementation; JP practitioners clone/audit grill-me from his repo before install.
URLs: x.com/mattpocockuk/status/2076949051610640813 · x.com/GuiBibeau/status/2076801729534845161 · x.com/noppezeirishi/status/2076950498876555356 · x.com/paulpitchford/status/2076776643310370842
Threatens: hand-curated personal skill trees with no packaging/version/promote-rollback story.
5. Tool-catalog ceiling — less MCP always-on, more on-demand
What's emerging: ~50 tools is a hard failure mode (wrong tool selection); practitioners standardizing on 4–6 MCP servers, least-privilege tokens, and on-demand tool discovery instead of full catalogs.
Evidence: Same-day recaps of Tech-With-Tim guidance: “Installing every Claude plugin can make Claude Code worse… once Claude sees around 50 tools, it starts picking the wrong ones”; Composio-style “find tool when needed” called out as the fix.
URLs: x.com/theSethian/status/2076971119899197445
Threatens: “MCP everything” / big plugin marketplaces as default. Next category: tool routers / lazy tool catalogs (not more servers).
6. Worktree-as-default multi-agent isolation
What's emerging: Git worktrees moved from “advanced tip” to default unit of agent concurrency (Claude on feature A, Codex on B, human on C; PR-review agents checkout worktrees hourly).
Evidence: Independent JP practitioner adoption post; Claude Code --worktree tutorial series framing “temporary construction site”; live PR-review routine: hourly scan → worktree → run app → comment.
URLs: x.com/kaigooddev/status/2076494363904049386 · x.com/kawai_design/status/2076788762894614768 · x.com/aks0510/status/2076983822550512050
Replaces: multi-session in one dirty working tree / branch-hopping.
7. Cross-harness dynamic workflows (Fable orchestrates non-Claude workers)
What's emerging: Not just multi-model chat — dynamic workflows where a Claude/Fable conductor spawns worktree’d foreign agents (herdr panes, pi agents on GPT-5.6-sol, headless Codex) and merges reviews; OSS “wiff” claims same pattern across harnesses.
Evidence: Live demo posts of Fable → herdr/pi/GPT-5.6 workers with worktrees; “tell Fable to spawn dynamic workflows”; herdr+pi+small custom skill set described as preferred daily stack; Claude→Codex control plugins in the wild.
URLs: x.com/oscabriel/status/2076783677552652456 · x.com/maxedapps/status/2076944322235089266 · x.com/xxxxoooxoxo/status/2076857108113559620 · x.com/PhilShteuck/status/2076783845832065197 · x.com/mvanhorn/status/2076372405388476789
Replaces: single-vendor subagents as the only fan-out surface. (You already run this — the shift is mainstream copy + productized dynamic workflows.)
8. Harness > model, code as substrate — but “general harness loses to specialized”
What's emerging: Research+practitioner consensus that harness is the product layer; Lilian Weng’s harness essay is the shared reference; Nirant’s read: code won for deterministic context engineering, and general harnesses (Claude Code, Codex) lose to specialized ones in production (frontier models for dev, cost/latency models for prod).
Evidence: Weng post defines harness as OS-like (workflow, eval, permissions, FS memory, subagents); Nirant’s take on that review (~120 likes, ~21k views, high bookmark rate for technical posts).
URLs: lilianweng.github.io/posts/2026-07-04-harness/ · x.com/NirantK/status/2075829950352068690
Replaces: “swap to a smarter model first” as the first remediation; challenges pure general-purpose multi-agent sprawl.
Terminology shifts (compact)
| Winning / rising | Contested / late-adopter | Dying as serious frame |
|---|---|---|
| Harness engineering (Weng, Nirant, ScottShapiroUXD pattern) | Loop engineering (Gergely skeptic + farm spam) | Vibe coding as competence claim (now slur / “incompetence promoters”) — x.com/annopost/status/2076700856494276803 |
| Agentic engineering (Karpathy line still recycled) | Agentic coding | Prompt engineering as the skill |
| Events engineering / event-triggered agents | ACE → proposed ADE (Agentic Development Environment) — thin trial balloon x.com/UtkarshUsername/status/2076760192453353865 | |
| Plugins (installable skill packs) | “AI employee” marketing language |
New voices (not on your ledger)
| Handle | Receipt | What ledger misses |
|---|---|---|
| @mattpocockuk | x.com/mattpocockuk/status/2076596585069174925 · plugin x.com/mattpocockuk/status/2076949051610640813 | Code-as-env doctrine, grill-me/skills packaging, plugin submission — highest-signal operational voice this window |
| @kenn | x.com/kenn/status/2076672785129353539 | Ambiguity-based routing: unknown → Fable / known → Codex; opportunity-cost defaults (Fable Low + Codex Sol Light) x.com/kenn/status/2076818001761022390 |
| @NirantK | x.com/NirantK/status/2075829950352068690 | Harness research distillation; specialized-vs-general harness thesis |
| @jpschroeder | x.com/jpschroeder/status/2076520407486628253 | Hard counter-signal: skills often hurt frontier models; MCP vs script directory |
| @koder0x | x.com/koder0x/status/2075825401058726249 | Concrete daily multi-agent loop + session-reset verify (low fame, high operational density) |
| @maxedapps | x.com/maxedapps/status/2076944322235089266 | herdr + pi + minimal skill set as preferred multi-agent UX |
Skip: Continuous-Claude / course-farm accounts (launch packaging, not unprompted adoption).
Ledger proposals
Add
- @mattpocockuk — currently the highest-yield operational signal outside your 20; skills and anti-skill-bloat.
- @kenn — best non-EN routing heuristics for Fable/Codex seats; fills JP power-user gap.
- @NirantK — harness-layer research→practice bridge your product/ship voices underweight.
Drop / demote (only if you must stay ≤20)
- Prefer demoting a recency-low, launch-only seat over any of karpathy / steipete / simonw / Gergely — no drop justified purely from this 84h window without your ledger yield stats.
- If forced: drop the weakest non-operational handle on your sheet (this scan does not name one with confidence).
Watch (don’t add yet)
@jpschroeder— useful adversarial seat; too sparse alone for a permanent slot.@koder0x— one excellent post ≠ ledger permanence; re-check in 2 weeks.
No wholesale list rewrite. Your conductor+peer-seat shape is confirmed mainstreaming, not threatened — the real drift risks are word rot (“loop engineering”), skill/plugin bloat, and always-on tool catalogs.