frontier

Frontier scan

Aug 3, 2026

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


1. Sol / Luna / Terra role routing as the default Codex architecture

What's emerging: Orchestrator (Sol) freezes a work order → cheap/fast executor (Luna Max) implements in scope → independent read-only auditor (Sol) grades Pass/Partial/Fail with ≤2 correction rounds; Terra only for heavy exploration.
Evidence it's adoption: Same pattern shows up independently as AGENTS.md policies, custom agents, paste-installers, and “Luna thread swarm” variants—not one vendor thread.
Replaces: One model does plan+code+review; vague “spawn subagents.”
URLs: x.com/sen_partho50106/status/2083652853596627057 · x.com/ChrisLyonsHFX/status/2083720534672495080 · x.com/KeisukeIshikawa/status/2083625335296299205 · x.com/hikariraina/status/2083788694675419468 · x.com/Voxyz_ai/status/2083545777264074782 · x.com/dungeonpunk/status/2083065133501198820 · x.com/verbscribe/status/2083996951096266858 · x.com/pocarles/status/2083948750708318271

2. One repository contract; thin per-harness adapters

What's emerging: Canonical AGENTS.md (install, architecture invariants, verification, forbidden changes, definition of done) + thin CLAUDE.md / Codex config that import it—not twin instruction piles.
Evidence: Long-form practitioner writeup this window + operators seeing Codex itself treat AGENTS.md as behavior owner; multi-agent dual-truth called out as silent drift.
Replaces: Parallel CLAUDE.md vs AGENTS.md “cultures,” per-agent private truth.
URLs: x.com/talwar_divyam/status/2083945120773341494 · x.com/i/article/2083617291175067648 · x.com/SearchDataEng/status/2083591826888413326 · x.com/talwar_divyam/status/2083993778873061556

3. Skill/MCP maximalism → prune + determinize

What's emerging: Power users delete overlapping skills; re-run models bare; push repeat work into scripts/hooks/CI; use MCP to discover, then freeze as deterministic paths.
Evidence: WorkOS talk recirculated as “deleted 95% of skills, harness did the work”; Boris “every 6 months delete CLAUDE.md/skills/hooks” still circulating as operating doctrine; collision complaints on skill piles.
Replaces: Skills-as-hoarding; “more MCP = more power.”
URLs: x.com/hhg2288/status/2083661185144983731 ·

· x.com/learnaifaster/status/2083829468842602665 · x.com/AlemTuzlak/status/2083975003884978487 · x.com/gregbarbosa/status/2082850592834142624

4. Job-earned multi-seat portfolio (not monogamy)

What's emerging: Capability is close enough that serious users route jobs (plan / implement / debug / autonomous / daily) across Claude Code, Codex, Hermes, Grok Build, etc., with harness transparency and cost as differentiators.
Evidence: Same-workflow multi-tool bakeoff framing + operator posts treating Claude-as-conductor / Codex-as-redline as normal, not exotic.
Replaces: “Pick the best coding agent” identity.
URLs: x.com/tonysimons_/status/2083601654247784701 · x.com/DAssetBuzz/status/2083255564574904780 · x.com/Nobullshit/status/2082905392145641977 · firecrawl.dev/blog/best-ai-coding-agents · mightybot.ai/blog/coding-ai-agents-for-accelerating-engineer

5. Company multiplayer harness (org OS, not personal CLI)

What's emerging: Category crystallizing: per-person isolation + shared rooms/skills/crons/memory, model-agnostic core (Codex/Claude/Pi/OpenCode), Slack+web—company work surface, not “another coding agent.”
Evidence: YC open-sourced QM they claim to run across accounting/legal/events/eng; real MIT repo with both AGENTS.md + CLAUDE.md, skills-seed, admin policy—treat as real product, discount pure launch theater.
Replaces / threatens: Personal Hermes/OpenClaw as the org default; single-user terminal harness as the only control plane.
URLs: x.com/ycombinator/status/2083243960684908768 · github.com/yc-software/qm · x.com/stretchcloud/status/2083290352404812263

6. “Harness engineering” > prompting as the craft name

What's emerging: The scarce skill is loop + tools + permissions + compaction + repair layers—not clever prompts; open-model gaps framed as contract/harness bugs.
Evidence: Progressive open harness curriculum posts; Command Code “validate-then-repair” tool-input work; provider-neutral harness skill repos/articles this window.
Replaces: “Prompt engineering” / “vibe coding” as the serious self-label.
URLs: x.com/DanKornas/status/2083332127546851776 · x.com/MrAhmadAwais/status/2083993061810651505 · github.com/DenisSergeevitch/agents-best-practices · medium.com/@tort_mario/ai-agent-best-practices-production-re

7. MCP demoted to thin I/O; skills = process; scripts win the hot path

What's emerging: Stack language settling as: ACP/editor surface → hooks/skills/rules → MCP tools; practitioners say people already swap chatty MCP for skills/scripts where possible.
Evidence: Independent stack posts + enterprise “MCP vs skills solve different problems” writeups; not “MCP is dead,” but “MCP alone is not the product.”
Replaces: MCP-as-the-whole-agent-platform.
URLs: x.com/cdbattags/status/2083249238708760800 · x.com/humntech/status/2083175115458744615 · developers.redhat.com/articles/2026/05/25/mcp-servers-vs-ski · x.com/gregbarbosa/status/2082850592834142624

8. Interview-the-operator workspace bootstrap (system-first)

What's emerging: High-signal operational move: inventory every daily tool (API/MCP/CLI) → plan-mode context collection interview → generate your structure/.mds—not copy a celebrity CLAUDE.md.
Evidence: One post this window hit ~440 likes / 711 bookmarks with pure how-I-work content (not product launch).
Replaces: Generic starter packs / prompt-first onboarding.
URL: x.com/EXM7777/status/2083270723649654851


Terminology shifts (this window)

Winning Dying / demoted
harness engineering, work order, repository contract prompt-first, dual instruction truth
Sol / Luna / Terra (role names as ops vocabulary) “just spawn a subagent” without ownership
company multiplayer harness / org agent OS personal coding agent as only unit
skill prune / scripts-first skills maximalism, MCP-everything
job-earned multi-seat routing monogamy / “best agent”

New voices (not on your ledger)

Handle Receipt What ledger misses
@talwar_divyam x.com/talwar_divyam/status/2083945120773341494 Cross-harness instruction architecture; Codex 32 KiB layered load; executable invariants over prose
@tonysimons_ x.com/tonysimons_/status/2083601654247784701 Hermes multi-agent ownership (credentials top, MCP inheritance off, skill least-privilege); cross-tool bakeoffs including Grok Build
@MrAhmadAwais / @CommandCodeAI x.com/MrAhmadAwais/status/2083993061810651505 Open-model harness craft: tool-input repair, schema design as “capability”
@DanKornas x.com/DanKornas/status/2083332127546851776 Progressive harness teaching (loop → tools → perms → memory → teams/worktrees/MCP)
@EXM7777 x.com/EXM7777/status/2083270723649654851 Operator-interview → personalized multi-tool workspace (Claude Code + Codex)
@hikariraina x.com/hikariraina/status/2083788694675419468 Failure receipts / narrow Luna worker contracts in production agent ops
@hhg2288 x.com/hhg2288/status/2083661185144983731 Harness-over-skills doctrine; explicit-state machines vs skill piles
@dungeonpunk x.com/dungeonpunk/status/2083065133501198820 Sol planner + Luna implementer skill as usage-stretch pattern

Ledger proposals

Add

  • @talwar_divyam — only clear multi-harness contract voice this window; matches your Claude+Codex+Grok seat risk.
  • @tonysimons_ — Hermes/Codex operational depth + ownership model your ledger (Claude-heavy) underweights.
  • @MrAhmadAwais — harness-as-craft / open-model tool repair; complements Cherny/product voices.

Watch (don’t add yet)

  • @EXM7777, @DanKornas, @hikariraina — useful ops signal; still lower consistency / engagement bar than ledger standards.
  • YC/QM — product category signal, not a person voice.

Drop

  • no drops — nothing on the list went silent or contradicted this window; risk is missing Sol/Luna + contract vocabulary, not obsolete handles.

Operate-now implications (for your setup)

  1. Freeze Sol-style work orders for Codex/Luna seats; cap correction loops.
  2. Make AGENTS.md canonical; CLAUDE.md only imports + Claude-only deltas.
  3. Schedule a skill/MCP prune against Opus 5 bare baseline (Cherny/WorkOS doctrine).
  4. Track QM / multiplayer harness as category—not a toy—if you ever need org/shared agent surface beyond personal conductor loops.