Frontier scan

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

1. Orchestrator never implements — main session is conductor only

What's emerging: Top session dispatches + grades only; implementers and reviewers are fresh windows; Blue/Red/synthesis review so the writer never grades own work. Replaces one long chat that plans and codes until context rot kills the plan.
Evidence: Independent builder write-up of hour-2 rot when main session implemented, with concrete conductor checklist (not a product launch).
x.com/vik_develops/status/2082462087536628189 · reply measuring replan cost: x.com/vukayardstick/status/2082492361548960173

2. Stack taxonomy winning: harness ⊃ graph ⊃ loop (loop is no longer the top noun)

What's emerging: Failure debugging is framed as “which layer broke?” — loop = retries/budgets; graph = topology/handoffs; harness = permissions/sandbox/evals. Replaces organizing primarily around “agentic loops.”
Evidence: Independent layered breakdown (not a single coordinated launch) with explicit loop⊂graph⊂harness; parallel “5-layer” stack posts in same window.
x.com/elune0x/status/2082493878238576748 · x.com/He1s_Sammy/status/2082496885273579854
Caveat: Many high-view “graph engineering” videos are recycled workshop spam — treat taxonomy as real, viral course clips as hype.

3. Plugins as the install unit (skills + MCP + config), not raw SKILL.md

What's emerging: Three-layer extensibility: MCP = live tools, skills = reasoning playbooks, plugins = distributable bundles. Install UX converging on /plugin marketplace add and npx skills add …. Threatens hand-copied skills and one-off MCP wiring.
Evidence: Practitioner layer split + live install chatter (npx skills add) + marketplace listing ops in-window.
x.com/stretchcloud/status/2082009915284209959 · x.com/nedeadins1de/status/2082547557280915600 · x.com/SoraiaDev/status/2082500548134056063

4. Claude Code as multi-model bus (official Codex-in-Claude + reverse plugins)

What's emerging: Cross-lab peer seats inside one host: OpenAI’s codex-plugin-cc (/codex:review|rescue|adversarial-review) plus reverse Claude-on-Codex and “Claude harness wraps other models.” Replaces tab-switching dual CLIs as the power default.
Evidence: Official repo + in-window rediscovery/adoption posts + reverse plugin shape.
github.com/openai/codex-plugin-cc · x.com/HeyAnjula/status/2081291179141018110 · github.com/sendbird/cc-plugin-codex · x.com/Harsh_Kapoor03/status/2082466423587401802

5. Agent Teams (peer mesh) vs nested subagents (tree)

What's emerging: Subagents = boss/worker isolation; Agent Teams = independent sessions that message each other and self-assign (still experimental). Replaces “more nesting depth” as the only multi-agent upgrade path.
Evidence: Extension chronology + experimental flag discourse + builder “design the team” framing this week.
pub.towardsai.net/claude-code-extensions-explained-skills-mc · x.com/AnandButani/status/2081764312486461909 · x.com/DaniilBuilds/status/2082225083507245488

6. Worktree isolation + external job orchestrators (not just in-process fan-out)

What's emerging: Parallel agents default to separate git worktrees; some builders run declarative multi-run orchestrators (incl. K8s) and “agent races” across harnesses. Replaces shared-cwd parallel sessions and pure in-process subagent hopes.
Evidence: Independent multirun ship + worktree parallel recipe traffic + multi-backend worktree runner.
x.com/kuberdenis/status/2082211321165078641 · x.com/mikenevermiss/status/2081984383838216308 · x.com/stretchcloud/status/2082502959204630713

7. Task state outside the chat; reviewer = rival model; skills over MCP for specialists

What's emerging: Orchestrator state lives in Kanban/signed events, not transcript; strongest prompt lines are prohibitions (“do not complete their work”); rival-provider reviewer on purpose; load specialists with skills not MCP (MCP slot can evict posting/tools). Replaces chat-as-source-of-truth and MCP-maxing specialists.
Evidence: Unprompted synthesis of public multi-agent setups with concrete anti-patterns.
x.com/bossriceshark/status/2082117932222693706 · Wiz/Atlas “durable advantage is the system”: x.com/LeonDerczynski/status/2082166433014845466

8. “Harness > model” is operational consensus; model-switch is a routing problem

What's emerging: Enterprise and security teams publicly say advantage is the system around the model; multi-model/multi-agent is assumed, not experimental; “model got dumber” reframed as context rot. Replaces model-chasing as primary lever.
Evidence: Independent cost/setup notes + multi-model ops posts + context-rot framing citing industry shift to context engineering.
x.com/triesai_co/status/2082548048169709662 · x.com/WiFiMoneyGuy/status/2082286433201447064 · x.com/piquopiquo/status/2081200917840347258 · x.com/bonduelleioat/status/2082529387941900308

Terminology shifts (this window)

New voices (not on your ledger)

Ledger proposals

Add

• vik_develops — highest-signal conductor-pattern ops this window
• stretchcloud — clean extensibility-layer language you can route decisions on
• kuberdenis — if you care about unattended multi-agent beyond one box
• vukayardstick — forces economic metrics (cost/PR) your current voices underweight
• bossriceshark — dense anti-pattern library for multi-agent setups

Drop / demote (optional)

• None mandatory from this sweep — ledger already covers core product voices (bcherny, steipete, theo, etc.). If forced trim: deprioritize pure product-announce accounts when their posts are mostly launch, not ops (none of your list clearly dead this window).

Watch only (not ledger yet)

• TheNoamLewis, Harsh_Kapoor03, eriks_b — useful, thinner sample
• Graph-engineering megaphone accounts (mikenevermiss, engagement recuts) — high volume, low originality

Bottom line for your setup: You are already on the right side of multi-seat harness + skills/hooks/MCP. Highest obsolescence risks are (1) treating loops as the top architecture word while the field promotes graph + harness layers, (2) main session still implementing, (3) skill/MCP without a plugin distribution layer, (4) subagent trees only while Agent Teams peer mesh matures.