Frontier scan

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

meme cascade (same diagram, many content accounts). Below keeps only items with independent practitioner receipts, product docs, or measured runs.

Ranked (max 8) — change-how-you-operate first

1. Hard law vs soft law: hooks + permissions over CLAUDE.md
What's emerging: Load-bearing constraints move out of model-read markdown into app-enforced hooks/permissions (format, block rm -rf, fail-closed on tests). Replaces “hope the model obeys CLAUDE.md.”
Evidence: Independent JP power users + trainers restating the same split (model-judged vs machine-guaranteed), not a single launch.
URLs: x.com/ClaudeCode_aca/status/2081230496986476699 · x.com/ryota110HR/status/2080820578303529212 · x.com/goro_claudecode/status/2080775036500164859

2. Minimal tool surface + progressive disclosure (search → execute)
What's emerging: Fat MCP catalogs are being collapsed to a tiny always-on set + deferred discovery; tools behave like skills. Threatens “install 40 MCP servers” as the default power setup.
Evidence: Hermes ships progressive disclosure behind one search-and-execute path (maintainer + field testers same window); Houston reports production cut to five tools.
URLs: x.com/Teknium/status/2081450522608107816 · x.com/ja818_/status/2081354345707348226 · x.com/sudoingX/status/2081453712149721457

3. Skills-over-MCP becomes a standards track (manual ships with the server)
What's emerging: MCP tools alone insufficient; skills ride MCP Resources so orchestration know-how ships with the connector. Next category after raw MCP.
Evidence: AAIF / Skills Over MCP WG write-up + experimental MCP ext + AAIF podcast push in-window.
URLs: aaif.io/blog/skills-over-mcp · github.com/modelcontextprotocol/experimental-ext-skills · x.com/AgenticAIFdn/status/2080745118424866992

4. Worktree fan-out as default parallelism (measured, not aspirational)
What's emerging: Multi-worktree / multi-branch Claude Code runs with merge-after-land; replaces single long chat.
Evidence: Practitioner posted measured run: 16 branches parallel, max 11 worktrees, ~2.7M tokens, 0 human code fixes, all merged same day.
URLs: x.com/Tebasaki_lab/status/2080884746671046686 · x.com/0xCodila/status/2081141384850108480 (amplifies worktree pattern; treat as secondary)

5. Cross-provider peer verify (Claude ⇄ Codex ⇄ Grok) as quality gate, not demo
What's emerging: Independent second-family review of Claude output; same-provider multi-agent is not enough.
Evidence: Unrelated practitioners describing the same ops pattern unprompted (VPS multi-seat, separate-model verify, 2–3 provider consensus).
URLs: x.com/cristianuibar/status/2080659898707193911 · x.com/pyra_m1d/status/2080315751122993154 · x.com/ShadesofSamsara/status/2081389160745566271

6. Claude Code artifacts = live shared work surface (and a new authz surface)
What's emerging: Session → private live page (PR walkthroughs, dashboards, timelines); team looks at one URL instead of “walk me through the agent.”
Evidence: Official Claude Code docs + independent security read on viewer-scoped MCP + one-click write auth gaps.
URLs: code.claude.com/docs/en/artifacts · support.claude.com/en/articles/9487310-what-are-artifacts-an · x.com/eh_hosseini/status/2081459645537767469

7. CLI+skills vs MCP: draw, not MCP-maximalism; axis is context + credentials
What's emerging: Power users stop treating MCP as the universal tool bus; ship CLI and MCP, or neither if tool surface is wrong.
Evidence: Substantive debate with tradeoffs (sandbox, tool discovery, cache, jq composition), not a product launch.
URLs: x.com/RhysSullivan/status/2080117243405635832 · x.com/andyhtran_/status/2080190880149844263 · blog.devgenius.io/saas-bootstrapping-in-2026-building-your-h (MCP “too heavy; CLI often wins day-to-day”)

8. Hermes Agent (Nous) as OSS harness contender for local + progressive tools
What's emerging: Open harness mindshare splitting: Hermes gaining for local endpoints, plugin ecosystem, MCP progressive disclosure; OpenClaw called out as heavier/failing local onboarding by field testers.
Evidence: Maintainer shipping tool scaling + independent local-agent runs (Bonsai 27B + Hermes on 8GB), not just announcement.
URLs: x.com/Teknium/status/2081450522608107816 · x.com/sudoingX/status/2081453712149721457 · x.com/Da7_Tech/status/2081419564051472840

Terminology shifts (window signal)

| | progressive disclosure / tool search | dump-all-tools into context | x.com/Teknium/status/2081450522608107816 | | skills + MCP resources | MCP-as-complete-solution | aaif.io/blog/skills-over-mcp | | graph / fleet / worktree fan-out (code-orchestrated) | single-agent chat loop as architecture | x.com/Tebasaki_lab/status/2080884746671046686 | | “graph engineering” as buzzword | ⚠️ mostly hype cascade unless state transitions + verifier edges are real | x.com/sidelined_cap/status/2081458477160124643 · x.com/ankit_appy/status/2081456100814332317 |

Your words check: agentic coding / harnesses / subagents / skills / loops are not dying — they are layering up. Risk is soft skill-prose without hooks, MCP sprawl, and same-provider multi-agent pretending to be peer review.

New voices (not on your 20-handle ledger)

Skipped as ledger candidates: content farms recycling the 5-layer poster (@free_ai_guides, @0xfuckpoverty, ECC “16k stars” recap bots) — amplification ≠ operational insight.

Tool-category shifts

Ledger proposals

Add

• @techgirl1908 — standards pulse on skills×MCP (your stack’s next protocol seam).
• @Teknium — Hermes tool-disclosure + open harness trajectory (Grok/local peer lane).
• @Tebasaki_lab — rare measured worktree-fanout receipts.
• @eh_hosseini — agent authorization / artifact blast-radius (your multi-seat risk).

Watch (don’t add yet)

• @ClaudeCode_aca — high-signal JP ops, but brand/edu account; sample 2 more weeks.
• @RhysSullivan / @ja818_ — one strong post each this window; need consistency.

Drop

• no changes among the existing 20 — all still active or definitional; none went silent or pure-hype this window.

Ignore noise

• Don’t promote “graph engineering” influencers without state-machine + verifier-edge receipts.
• Mega skill packs (everything-claude-code style) = optional catalog, not architecture.