Logos52
wiki / Systems / AI and Agentic Systems / Current Agentic LLM Stack

Current Agentic LLM Stack

reference updated 2026-09-01

Current Agentic LLM Stack

The models and agents doing agentic work across the projects, as of 1 September 2026. Grok 4.6 writes vault prose and executes. Cowork is research on request, not the default writer. The only local models left are audio instruments.

Current Stack

LayerTool / ModelPurpose
Primary writerGrok Build (grok-4.6; grok-n1 / grok-admin profiles)Wiki, journal, decisions, reports; final cut at the desk. Signed 1 September 2026
Research on requestClaude Cowork (Fable 5 / Opus 5)Lanes and structure when asked. Not the default writer. Fable’s late-August prose was a rewrite tax
Vault pipelineGrok Build calling 4.6Clean-context rewrite and cold read as fresh sessions, plus repo work on the vault
Coding agentGrok Build 1.0.5 (grok-4.6 as of 2026-08-12; grok-n1 / grok-admin profiles)Execution on the Mac’s real files and toolchain; separate profiles keep personal and work isolated
IDE seatCursor (Ultra active, free month via Heavy, expires 2026-09-12)Sitting in application code: tsumugu-core, tan, frontends; Tab and visual diffs. Not the vault’s author
Standing agentsGrok Bot (cloud, one shared computer)The always-on half — watching, fetching, filing; see Standing Research Agents
Local modelsQwen3-TTS + Whisper (MLX, Apple Silicon)Audio production only: TTS generates the voices, Whisper transcribes them back for QA
InstructionsAGENTS.md / CLAUDE.md per repoProject rules, loaded at session start

Division of Labor

Grok 4.6 carries vault writing and execution on the machine where the files live. Cowork still runs research lanes when asked. It does not take the default writer seat: late-August Fable prose cost extra tokens to make readable. Grok Bot carries the standing lanes on its own cloud computer, which holds only what is already public. Grok Bot shares the same plugins and connectors as the Cursor account this desk already has, so a Cursor user pays less setup cost than someone starting from nothing. That shared plugin surface is not a reason to put a private login on the shared computer. The local models are production instruments rather than agents: voice generation and its transcription check, both on Apple Silicon.

Everything runs on subscription plans or local hardware; nothing in the stack is pay-per-token. Claude Managed Agents, the Agent SDK, and the Messages API are that meter (tokens, and for Managed Agents $0.08 per running session-hour). Ruled off the roster 2026-08-28: “i don’t like anything with API.”

Routing by job

The by-job split, signed 1 September 2026 (Grok writes), with the 21 August seat assignment still in force for Cursor vs Build (the seat assignment). The 15 August ranking that put Fable on wiki prose is retired for routing and kept as history (the ranking).

JobSeat
Anything that has to read well — wiki prose, journal, decisions, reportsGrok 4.6
Research banks, batch authoring, regen, lint, vault scriptsGrok Build calling 4.6
tsumugu-core reader, tan Swift app, dashboard and frontend workCursor
Standing duty — one duty, packet only, public material onlyGrok Bot
Overnight code on a non-wiki repo, PR as the handoffCursor Cloud Agent, approved through Bot

One writer per tree — two agents editing the same files failed on 12 June and the roster was cut within the hour. Grok Bot never writes into wiki/; it drops into raw/inbox/ and Build compiles with the owner looking.

The ideal week — signed 2026-08-26

The rows above are ruled; the four below were proposed from the 26 August research pass and signed by the owner the same day:

  • The Ultra month is spent only where Build is weak — Tab, visual review, the Xcode project, one overnight Cloud Agent — and closes by its own written condition: Tab unused, Cloud Agents unused, Ultra at 0% means the month ends and the stack page does not change. Ultra is active and expires 12 September 2026.
  • The writing pipeline’s rewrite pass and cold read run as fresh Grok sessions holding only the draft and the prompt, never inside the writer’s session.
  • Effort dials are set per job, not left on default: xhigh in Build for hidden bugs and high-value diagnosis, lower effort in Cowork for routine orchestration, since every harness now documents the dial.
  • The two unrun experiments — Arm D (Fable pinned in Cursor, one scored bank) and the one overnight Cloud Agent — either run before the Ultra month ends on 12 September or are dropped as decided, not left open.

What’s Gone

Hermes 3 via Ollama — the May version of this page named it the primary interface for wiki work. It was retired within weeks and no local chat or agent model replaced it: local inference at useful agent quality cost more in speed and reliability than the subscription agents charge, and the privacy case it was meant to serve is handled instead by keeping private material out of cloud-agent reach entirely.

Evolution

  • Early 2026 — Grok for all knowledge-base work.
  • May 2026 — local-model experiment: Hermes 3 8B via Ollama as the wiki’s agent interface. Retired within weeks.
  • June 2026 — Claude Cowork becomes the primary agent; vault prose moves under the writing standards.
  • August 2026 — Grok Bot (beta 2026-08-11) adds the standing half; local models narrow to audio production. Bot access widens on 2026-08-21 to SuperGrok Plus, Cursor Pro+, and Cursor Teams. Cursor.app lands on the Mac 2026-08-26 with the Ultra month still unspent. 13–15 August: Fable wins a scored wiki bakeoff and takes the writer seat.
  • 1 September 2026 — Writer seat moves to Grok 4.6. Fable’s late-August prose was a token tax. Grok writes.

Sources