frontier

Frontier scan

Jul 23, 2026

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


Top 8 (ops-change rank)

1. Harness redesign goal = “locally in-distribution” (LID), not more tools

What's emerging: Treat the harness as the compositional generalizer: context offloading + programmatic sub-calls so root trajectories stay token-similar across lengths/domains; naive Claude Code/Codex-style “append everything” is called out as OOD/context-rot.
Evidence: Independent MIT/research writeup with experiments (30B RLM vs base Transformer, 8–32× length transfer); not a product launch.
URLs: x.com/a1zhang/status/2079203524395573442 · alexzhang13.github.io/blog/2026/harness/
Replaces: “Bigger context + more MCP tools” as the main scaling lever.

2. Empirical reality check: context files carry the stack; skills/subagents are thin

What's emerging: Across ~2.8k repos, Context Files dominate (often sole config); AGENTS.md is the interop standard; Skills/Subagents usually 1–2 artifacts and mostly static instructions, not executables; no spend-cap config surface in the wild.
Evidence: Peer-style empirical study + public dataset (not vendor marketing); practitioners restate the same finding unprompted.
URLs: assets.empirical-software.engineering/pdf/aiware26-agents-co · se-uhd.de/ai-config-dataset/ · x.com/enginenerdx/status/2079678693396431353
Replaces: Assumption that “skills + subagents + plugins” is where power users actually put weight.

3. Separate builder vs verifier is becoming table stakes (not optional review)

What's emerging: One sub-agent builds; a different agent/session verifies; self-grade is treated as systematically soft. Paired with /goal-style exit conditions so AFK loops don’t poll for permission.
Evidence: Anthropic Claude Code workshop recirculating with concrete op recipes; independent multi-provider users already wire Fable orchestrate → Opus code → Codex review loops.
URLs: x.com/Serantych/status/2078836146893619374 · x.com/Sebrock1972/status/2079998557839655281
Replaces: Single-agent “review your own PR” as the quality gate.

4. Spec → tickets → AFK agent is productizing (skills as installable discipline)

What's emerging: Grill-to-spec, ticket decomposition, TDD/bug loops as shared skill packs pluggable into Claude Code and Codex plugins; explicit “AFK agent” composition.
Evidence: Author shipping public skills + plugin packaging (Claude plugin approved; Codex plugin in flight) with operational thread replies, not a one-day hype blast.
URLs: x.com/mattpocockuk/status/2079318617041035618 · x.com/mattpocockuk/status/2080000544102375889 · x.com/Yamik1shi/status/2079563118133354749
Replaces: Chat-shaped “vibe a feature” as the default agent session shape.

5. Structural codebase memory is becoming a real category (next “MCP” for coding agents)

What's emerging: Persistent code knowledge graphs (call chains, routes, impact) queried over MCP instead of re-grepping; claimed ~10× fewer tokens / ~2× fewer tool calls on structural Qs.
Evidence: Open binary + paper + multi-client install path; independent roundups treat “agent memory” as structural retrieval, not chat logs; practitioners wiring skills around it on Grok Build.
URLs: github.com/DeusData/codebase-memory-mcp · arxiv.org/abs/2603.27277 · x.com/milvusio/status/2079944561393037580 · x.com/PixelShipper/status/2079949165568664046
Threatens: Session-amnesia fixed only by longer CLAUDE.md + grep/read loops.

6. Model routing is productizing; power users already hand-roll it in AGENTS.md / subagent model pins

What's emerging: Task-level model selection (frontier planner / cheap executor / specialist reviewer) as default economics; Cursor ships Router (~60% cost claim).
Evidence: Vendor productization plus independent practitioners describing Fable orchestrator + cheaper subagents / multi-CLI peers without waiting for Cursor.
URLs: x.com/cursor_ai/status/2079993729532989500 · x.com/DavidMansaray/status/2080003145166393582 · x.com/QuaintTransfer/status/2079985869830672429 · x.com/_nickanthony/status/2079314535526088811
Replaces: “Always Opus/Fable for everything” as the sophisticated setup.

7. Enforcement split is solidifying: hooks = hard; CLAUDE.md/AGENTS.md = advisory

What's emerging: Shared doctrine that prose memory is best-effort; hooks (and policy gates) are the only reliable rule layer; keep CLAUDE.md short.
Evidence: Multiple independent operators saying the same thing (not one coordinated launch); cross-harness caveat also noted (hooks not portable → vendor lock).
URLs: x.com/MarsBased/status/2079461257967513666 · x.com/evoclock/status/2079517345131659703 · x.com/itsmistermoon/status/2079458452108144985 · x.com/Divyyanshishrma/status/2079163727379931186
Dying term: “Rules/memory will make the agent obey.”

8. Multi-CLI peer seats + worktree isolation = default multiplayer architecture

What's emerging: Parallel Claude/Codex/Grok sessions on git worktrees; cross-provider review as policy; Grok Build auto-loading Claude skills/hooks/MCP as migration path.
Evidence: Independent builders describing worktree+lease setups and Claude↔Codex mutual-consult rules; Grok Build compatibility claims from users (not only @grok).
URLs: x.com/kulekci/status/2079796046599721013 · x.com/rcmisk/status/2079325625739821475 · x.com/PerfectlyFriedB/status/2080012806708179349 · x.com/poweroverthink/status/2079160608537116892 · x.com/bradshannon/status/2078658072629715330
Replaces: One mega-session / one provider as the “serious” setup.


Terminology shifts (this window)

Winning Dying / demoted Receipt
Harness engineering prompt engineering as the craft x.com/vishalsingh2972/status/2080007924081213950 · alexzhang13.github.io/blog/2026/harness/
Loop engineering (design the self-talk-to-done loop) “just prompt better” x.com/dabsurdweb3/status/2080007682715529337
AGENTS.md (interop context file) tool-private only (CLAUDE.md alone) assets.empirical-software.engineering/pdf/aiware26-agents-co
AFK agent / exit-conditioned /goal chatty interactive agent x.com/mattpocockuk/status/2080000544102375889 · x.com/Serantych/status/2078836146893619374
Structural / graph memory “memory” = embed chat history x.com/milvusio/status/2079944561393037580
Model router single-model default x.com/cursor_ai/status/2079993729532989500

New voices (not on your ledger)

Handle Receipt What they cover that ledger misses
@mattpocockuk x.com/mattpocockuk/status/2079318617041035618 Installable discipline skills (spec/tickets/TDD/review), AFK composition, Claude+Codex plugin packaging
@a1zhang x.com/a1zhang/status/2079203524395573442 Formal harness theory (LID, Hi/Q trajectories, why append-history harnesses fail)
@drummatick x.com/drummatick/status/2079096051810857298 When not to custom-harness: free oracle (tests) vs irreversible/no-oracle work
@enginenerdx x.com/enginenerdx/status/2079678693396431353 Adoption telemetry → “boring context file wins”; spend-cap gap callout
@lateinteraction alexzhang13.github.io/blog/2026/harness/ (co-author credit in thread) RLM / retrieval-harness research co-signal with a1zhang

Skipped as hype/low-ops: recycled 4-layer Claude carousels, crypto agent-builder launches, single-repo “I built a harness” teasers.


Ledger proposals

Add

  • @mattpocockuk — highest-signal operational skill author shipping AFK-ready packs this window
  • @a1zhang — only clean formal upgrade to what “harness” means; will reframe your architecture reviews

Optional add (if you expand past 20)

  • @drummatick — crisp decision rule for custom harness vs stock Claude/Codex

Drop / demote (only if slot-constrained)

  • Prefer demoting pure industry-news / hiring-macro seats over operator seats when you need room (e.g. @GergelyOrosz if forced — less agent-harness ops density this window).
  • No drop required if 20 stays fixed: your list already covers Anthropic/Claude Code, Codex ecosystem, and systems voices; the miss is skill-pack + formal harness theory.

Watch (not yet ledger)


Skeptical rejects (not ranked)

  • Mass-duplicated “Claude = 4 layers” carousels (same graphic, low original practice).
  • Oxford “$4.8M graph memory 4×” style engagement bait without primary link quality.
  • On-chain “Agent Builder / spend cap” product spam.
  • Cursor Router as the story alone — real product, but practitioners already hand-routed via subagent model pins / multi-CLI.

Ops implication in one line: stop adding clever skills until AGENTS.md + hard hooks + separate verifier + LID-style context offloading + structural memory are load-bearing; skills without those are decoration (data agrees).