backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 84h
Window: ~2026-08-06 → 2026-08-10 01:00 IST (84h) · all-X + web · adoption over announcement
Ranked (max 8) — would change how you operate
Agent Plugins 1.0 as the portable packaging layer (next “MCP” for distribution) — Multi-vendor ship (Vercel/AWS/Cursor/GitHub/MS/OpenAI; Google joined TSC) + independent implementers (Context7 portable plugin, Hermes portable install) within days; one
plugin.json+skills/+mcp.jsonreplaces forked per-client plugins. Threatens client-locked Claude Code plugins / hand-maintained skill trees as the shipping unit.
vercel.com/blog/introducing-agent-plugins · x.com/vercel/status/2085398393179005223 · upstash.com/blog/context7-portable-agent-plugin · x.com/enesakar/status/2086498636687659095 · x.com/_vmlops/status/2086484035669504105Graph engineering (DAG/orchestrator-workers) displacing pure sequential loops as the power-user unit of work — Cherny-sourced framing (“building graphs and loops… in 3+ months we’ll all be doing graph engineering”) is being re-taught and operationalized as parallel subagent nodes with edges/state; independent pushback that it’s partly a rename of existing patterns is real, so adopt the structure not the slogan. Replaces single-threaded agentic loops / “one chat until done.”
x.com/0xMorlex/status/2086059854704365863 · x.com/CoinSh0t/status/2086509203187745255 · x.com/0xtriplug/status/2086532581705543687 · x.com/eddieor/status/2086426462949621839Skills (procedure files) winning default extension slot; MCP demoted to live-tool niche — Practitioners + Anthropic-course framing treat skills as the first reach; MCP called out as heavy/slow/token-bloated for day-to-day (CLI+README / skill playbooks preferred; MCP kept for Slack/DB/auth-shaped tools). Replaces “MCP everything” and prompt-only specialization.
x.com/Andrew_FG_AI/status/2085715939144691999 · blog.devgenius.io/saas-bootstrapping-in-2026-building-your-h · x.com/Keldrik/status/2086461884438573290 · x.com/VaibhavSisinty/status/2086429819735310420Harness↔model decoupling (“freedom of harness”) + multi-model pipelines as normal ops — OpenAI/Anthropic public exchange + practitioners running Fable plan → Sol adversarial → cheap Luna implementers in worktrees → Opus/Sol final review; Codex models inside CC harness treated as feature not heresy. Replaces single-vendor lock-in and “one model seat does everything.”
x.com/thsottiaux/status/2086186284528374095 · x.com/thsottiaux/status/2086188036493344823 · x.com/essepl/status/2086522402469748879 · x.com/0xAndros/status/2085538522262561202Cross-session agent messaging = multi-terminal teams without paste — Claude Code sessions can
SendMessage/ discover each other (macOS/Linux); independent builders writing team playbooks (migration finishes → waiting session pings). Replaces copy-paste handoffs and isolated tab silos for multi-agent local runs.
x.com/PrajwalTomar_/status/2086496738916040831 · x.com/Austen/status/2086417994461139067 · x.com/axyojp/status/2086481481027965068Stop conditions / steady-state as the missing loop primitive (esp. unattended) — Real product loop: 19 cycles, 44 commits, then self-cancels cron after two no-safe-work cycles; separate ops lesson that “stop when done” is not a stop condition without done-check + file allowlist. Replaces unbounded cron loops and “just keep improving.”
x.com/staskulesh/status/2085402240479596995 · x.com/cozybearlog/status/2086380353259020576Long-session runtime (hours) + server-side compaction as table stakes — Independent report of Codex staying on target 6–7h with test/env feedback; Hermes shipping OpenAI Responses native compaction for GPT-5.6 long runs. Replaces babysat short sessions and naive local transcript growth.
x.com/Dan_Jeffries1/status/2086535190520660065 · x.com/tonysimons_/status/2086436348659929376 · x.com/rauchg/status/2086513316265181213Agent infra is not chat: cache death across turns/model switches is the cost driver — Copilot production-scale paper (13M sessions / 95T tokens) + Hausfather-style metering (agent loops, cache-dominated energy) make session-aware scheduling, avoid mid-run model flips, and compaction policy operational. Replaces chat-era token mental models and casual multi-model mid-task switches.
arxiv.org/html/2608.00101v1 · cryptobriefing.com/microsoft-github-copilot-sessions-researc · x.com/rehen sina/status/2086425766741594376 · x.com/EvanKirstel/status/2086111441266889073
Terminology shifts (this window)
| Winning / rising | Dying / demoted | Evidence |
|---|---|---|
| graph engineering (+ still loop / harness engineering) | prompt engineering as the job | Cherny quote cycle; Ng “demoted” framing — x.com/0xMorlex/status/2086059854704365863 · x.com/cashhheth/status/2086520553037255011 |
| skills / SKILL.md / behavior layer | “just MCP it” / tool-server-first | x.com/VaibhavSisinty/status/2086429819735310420 · blog.devgenius.io/saas-bootstrapping-in-2026-building-your-h |
| Agent Plugins (portable package) | per-client plugin forks as the standard | vercel.com/blog/introducing-agent-plugins |
| memory + skills + hooks + subagents + plugins stack language | chatbot / single-prompt framing | x.com/VaibhavSisinty/status/2086429819735310420 · x.com/dan__rosenthal/status/2086165641594048875 |
| CLI for day-to-day tools | MCP as default for every tool | x.com/Keldrik/status/2086461884438573290 · blog.devgenius.io/saas-bootstrapping-in-2026-building-your-h |
Skeptical filter: “Graph engineering” is partly buzzword recycling of orchestrator-workers/parallelization (callout); treat Cherny’s practice (graphs+loops over one-shot prompts) as signal, LinkedIn courses as noise. Agent Plugins is multi-vendor real, but v1 deliberately omits hooks/subagents/commands — not a full harness replacement yet (Upstash limits).
New voices (not on your ledger)
| Handle | Receipt | What they cover that ledger misses |
|---|---|---|
| @mattpocockuk | github.com/mattpocock/skills · x.com/ze_ming94607/status/2085982077930090546 | Small composable eng skills (/grill-with-docs, /tdd, /handoff, /to-tickets) for Claude Code and Codex — process-as-skills, not hype architecture |
| @thsottiaux | x.com/thsottiaux/status/2086186284528374095 · x.com/thsottiaux/status/2086188036493344823 | Codex/OpenAI side of harness freedom, Sol-in-CC, limit resets — peer to bcherny you don’t currently track |
| @enesakar | x.com/enesakar/status/2086498636687659095 · upstash.com/blog/context7-portable-agent-plugin | Portable Agent Plugins packaging (file locations = contract); Context7 as multi-client skill+MCP unit |
| @staskulesh | x.com/staskulesh/status/2085402240479596995 | Unattended loop with stop/steady-state + production metrics (commits/tests/deploy), not just “run overnight” |
| @dan__rosenthal | x.com/dan__rosenthal/status/2086165641594048875 | Company-scale ops harness: process map → MCP/CLI → markdown brain → 50+ skills → hooks — non-code execution layer |
| @PrajwalTomar_ | x.com/PrajwalTomar_/status/2086496738916040831 | Agent-team playbooks for cross-session messaging (the new multi-terminal coordination pattern) |
| @tonysimons_ | x.com/tonysimons_/status/2086436348659929376 | Hermes/Codex long-session ops (native Responses compaction) — runtime knobs your ledger rarely covers |
Ledger proposals
Add
- @thsottiaux — OpenAI/Codex harness-policy peer to bcherny; freedom-of-harness is live doctrine for multi-seat setups
- @mattpocockuk — highest-signal skill-author voice for engineering fundamentals as installable skills
- @enesakar — early portable-plugin implementer; tracks the packaging standard you’ll ship against
- @staskulesh — stop-condition / unattended-loop receipts (directly relevant to scheduled agents)
Watch (don’t add yet — thin independent volume or narrow)
- @dan__rosenthal (strong ops stack post; confirm consistency)
- @PrajwalTomar_ (messaging playbooks; more productized content)
- @tonysimons_ (Hermes-specific; add if Hermes stays in your seat mix)
Drop
- no drops this window — none of the 20 went silent or pure-hype in a way that warrants culling; graph-engineering content farms are the false positives, not your existing handles.
Bottom line for your setup: The words still work (harness / skills / hooks / subagents), but the unit of shipping is shifting to Agent Plugins, the unit of work is shifting to graphs with stop conditions, and the default extension is skills+CLI first, MCP when live tools demand it. Highest leverage deltas this week: portable plugin packaging, multi-model plan/review/implement pipelines, cross-session messaging, and hard stop predicates on every scheduled loop.