backend grok-cli (SuperGrok Heavy sub, $0 marginal) · window last 84h
1. Commit-/pre-PR security becomes the real gate — PR review is drowning
What's emerging: At high cloud-agent PR volume, the human review checkpoint moves earlier (pre-commit / agent-CI loops / “proof packets”), not later.
Evidence it's adoption: Corridor reports ~95.5% of their PRs from cloud agents, review cycles ~15×, and a real near-leak fixed pre-commit; independent builders describe multi-agent CI-until-green and proof packets before PR.
Replaces: “PR review is the security/quality gate.”
URLs: x.com/corridor/status/2085079395937792178 · x.com/SaaiArora/status/2084405514629505246 · x.com/imrobertmine/status/2084703747343892774
2. Multiplayer / shared-room agents (teams, not private 1:1 sessions)
What's emerging: Agents as channel-visible coworkers with mailboxes, handoffs, and peer rooms (Claude Tag trajectory + native Agent Teams + multi-CLI shared rooms).
Evidence it's adoption: Live Agent Teams run with specific failure modes (task-list vs disk ownership drift, conflicting “agreed” ranges, silent out-of-scope decisions); infra voices argue SLDC must absorb swarm volume; product thesis is multiplayer, not more CLI tips.
Replaces: Solo conductor session as the only unit of work.
URLs: x.com/EverymansAI/status/2084364865955934314 · x.com/benyficent/status/2084786381918966236 · anthropic.com/news/introducing-claude-tag
3. Codex harness generalizes past “coding agent” → Work / knowledge-work OS
What's emerging: Same harness (subagents, browser, skills, plugins, scheduling, persistent cloud VM) is being productized as Work for non-dev knowledge work at ChatGPT scale.
Evidence it's adoption: Outside reconstruction (Aug 4) of Memory / proactivity / heartbeats / plugins+skills; ~10M Work+Codex users claimed; swyx frames continuous harness-engineering coverage as the beat.
Replaces: Treating “harness engineering” as a coding-only discipline.
URLs: latent.space/p/unpacking-chatgpt-work · x.com/swyx/status/2085073764551962960 · x.com/latentspacepod/status/2085072791087480947
4. Portable Agent Skills (open standard) win over per-tool prompt packs
What's emerging: SKILL.md + progressive disclosure as cross-harness contract (Claude Code, Codex .agents/skills, Copilot, Cursor…); vendors ship skills as plugins (e.g. Firebase → Codex).
Evidence it's adoption: Official agentskills.io + Codex docs usage notes from practitioners; Firebase GA skill install path for Codex; harness compatibility matrices treat .agents/skills/ as the broadest shared path.
Replaces: “CLAUDE.md-only / copy-paste prompt library” as the reuse layer.
URLs: agentskills.io/home · x.com/cv_usk/status/2084826978058707371 · x.com/Firebase/status/2085081703874228727 · codylindley.github.io/ai-harness-engineering-compatibility-m
5. Behavior evidence > green-diff skimming (walkthroughs, recordings, proof packets)
What's emerging: Review artifact shifts from “400 lines of green” to “here’s it running” + structured proof (commit/tree SHAs, tests, env).
Evidence it's adoption: Practitioners admit agent PRs aren’t really read; tools/workflows that attach walkthrough/recording or multi-agent proof packets are getting unprompted “this is the right direction” takes.
Replaces: Diff-only review as the human bottleneck.
URLs: x.com/HelloVyom/status/2085062971768897690 · x.com/imrobertmine/status/2084703747343892774
6. Worktree-isolated parallel subagents are becoming product default language
What's emerging: Fan-out into isolated worktrees (main copy untouched) + persistent background agents + local audit log — now marketed as the default SE agent architecture (Meta Muse Code beta; already hand-rolled by multi-CLI users).
Evidence it's adoption: Independent spawn-into-worktrees patterns (Hermes/Claude/Codex) before Meta’s post; Meta codifies the same shape as first-class.
Replaces: Single shared working tree multi-agent edits (and pure sequential solo agents).
Caveat: Muse Code itself is launch-day — architecture convergence is the signal, not Muse adoption yet.
URLs: x.com/finkd/status/2085080750034940201 · x.com/JamesMeadlock/status/2085074328228360700 · x.com/morpphhhaw/status/2085002027688460496
7. Hooks / runtime authority as the hard control plane (skills = soft procedure)
What's emerging: Clear layering: MCP = hands, skills = how/when procedures, hooks/runtime policy = must-run deterministic gates (block, test, scan, inject).
Evidence it's adoption: Practitioners teaching hooks as multi-handler lifecycle with block/allow return paths; security accounts reframe bottleneck as “what the agent can read/call/deploy/remember,” not model IQ.
Replaces: Hoping the model “remembers” CLAUDE.md rules under load.
URLs: x.com/mrizkimurtadha/status/2085034294502969592 · x.com/Hem_chandiran/status/2084151274971558191 · x.com/AgentGuard_AI/status/2084266644969500832
8. “Use AI to build the deterministic app, not to run every task”
What's emerging: Power users encode repeatable workflows into normal software/scripts after agents discover them; reserve agents for residual uncertainty.
Evidence it's adoption: Independent Medium + X framing against total agentization; pairs with hooks/skills as the middle layer (agent writes the gate, gate runs without LLM).
Replaces: Leaving every recurring chore as an agent loop.
URLs: x.com/alvincho/status/2084075559902994609 · medium.com/agentive-futures/use-ai-to-build-the-app-not-to-r
Terminology (84h pulse)
| Winning | Dying / demoted | Receipt |
|---|---|---|
| Harness engineering (still the serious frame) | “Just prompt better” | x.com/swyx/status/2085073764551962960 · latent.space/p/unpacking-chatgpt-work |
| ADLC (agentic dev lifecycle) | SDLC-as-usual + Copilot seats | x.com/fseixas/status/2085078814304993502 |
| Multiplayer / coworker agent | Private copilot chat | anthropic.com/news/introducing-claude-tag · x.com/benyficent/status/2084786381918966236 |
| Plugins (OpenAI: apps+skills+MCP tools) | Scattered “connectors / GPTs / Actions” | latent.space/p/unpacking-chatgpt-work |
| Commit-time / pre-PR security | PR-as-sole gate | x.com/corridor/status/2085079395937792178 |
| Behavior review / proof packet | Diff skimming | x.com/HelloVyom/status/2085062971768897690 |
| Agent Skills (portable standard) | Tool-specific prompt packs | agentskills.io/home |
Still noisy / not ready to rename around: pure “vibe coding” as serious practice (declining among power users; still spam-layer language).
New voices (not on your ledger)
| Handle | Receipt | What ledger misses |
|---|---|---|
| @shloked | x.com/swyx/status/2085073764551962960 · latent.space/p/unpacking-chatgpt-work | Lab harness reverse-engineering (Work/Codex/memory) at product-internals depth |
| @KentonVarda | x.com/KentonVarda/status/2084293600142135365 · x.com/KentonVarda/status/2085078529935556807 (quoted Cloudflare OS thread) | Sandbox-as-product (Gadgets/grains); agent-modifiable instances; pi-agent-core adoption signal |
| @corridor | x.com/corridor/status/2085079395937792178 | Commit-time agent security under extreme agent-PR ratios (ops numbers) |
| @benyficent | x.com/benyficent/status/2084786381918966236 | Cloud-agent / multiplayer pivot thesis (Codex history → Tag/infra implications) |
| @HelloVyom | x.com/HelloVyom/status/2085062971768897690 | Behavior-first agent PR review practice |
| @cv_usk | x.com/cv_usk/status/2084826978058707371 | Codex Skills progressive-disclosure ops notes (paths, implicit invoke, config) |
| @JamesMeadlock | x.com/JamesMeadlock/status/2085074328228360700 | Multi-CLI spawn into isolated worktrees (Hermes/CC/Codex) |
| @SaaiArora | x.com/SaaiArora/status/2084405514629505246 | Agents-as-CI (fix until green + auto-reply to review bots) |
Skip as net-new ledger seats this window: engagement-farm “Resource Bible” cloners; pure launch amplifiers of Muse Code.
Tool-category shifts
| Emerging category | Threatens | Evidence |
|---|---|---|
| Portable skills packs (agentskills.io) | Per-IDE rule files / prompt libraries | agentskills.io/home · x.com/Firebase/status/2085081703874228727 |
| Commit-time agent security scanners | Human-only PR security review | x.com/corridor/status/2085079395937792178 |
| Plugins (skills + MCP apps + templates) | Standalone MCP directories as the only integration story | latent.space/p/unpacking-chatgpt-work |
| Agent CI / proof-packet gates | “Hope the agent ran tests” + thin CI | x.com/SaaiArora/status/2084405514629505246 · x.com/imrobertmine/status/2084703747343892774 |
| Shared multiplayer rooms (Tag/Teams/peer CLIs) | Solo CLI as team collab surface | x.com/EverymansAI/status/2084364865955934314 · x.com/g00manoid/status/2084212417211371611 |
| Fine-grained sandbox “gadgets” (Cloudflare OS) | Unscoped local agent full-FS authority | x.com/KentonVarda/status/2085078529935556807 (thread root x.com/KentonVarda/status/2084990137180590572) |
MCP is not dying — it’s being demoted to plumbing under skills/plugins/hooks.
Ledger proposals
Add
- @shloked — best current public reverse-engineer of frontier harnesses (Work/Codex/memory).
- @KentonVarda — sandbox/security-shaped agent product reality, not tooling cosplay.
- @benyficent — cloud/multiplayer agent infra narrative with operational framing.
- @HelloVyom — review-surface shift (behavior evidence) you’ll feel as PR volume rises.
- @cv_usk (optional) — high-signal Codex skills ops if you want non-EN ledger diversity.
Drop / demote
- None forced from this 84h alone — your core set still owns the canonical narrative seats.
- Soft demote candidates only if capacity-bound: pure research/essay voices if you need more ops density (none of the listed names failed the window).
Watch (don’t add yet)
- @corridor — high signal but vendor; sample 2–3 more independent commit-time stories first.
- Muse Code — architecture confirmed; wait for unprompted multi-week usage reports before any Meta-adjacent voice seat.
Skeptical one-liner: most X “Claude Code stack” posts this window are recycled infographics; the load-bearing moves are earlier gates under agent volume, multiplayer shared state, skills portability, and Codex harness escaping the IDE into Work.