Logos52
wiki / Research / Grok Bot Field Packet 2026 08 15

Grok Bot Field Packet 2026-08-15

research updated 2026-08-31

Grok Bot Field Packet — 2026-08-15

Verdict: the pinned list confirms the 13 Aug bank on architecture and does not change the stack. Three mechanisms are new enough to absorb: official first-task shape (outcome / sources / constraints / deliverable / review point); Nate’s “a Bot owns a theme, not a task”; Teach-a-task holds on short recordings and goes brittle on long ones. Grocery, Gmail, Amazon, and spend-without-a-gate stay refused.

Novelty note. Entries checked: Practitioner Bank (Palmer essay + docs, 2026-08-13), A Field Tracker (assignment, no prior packet). This is the first Field packet. It differs by compiling the rest of the named-runner list, not by re-banking Palmer.

Pinned list: raw/sources/2026-08-15-grok-guides. Lane: A Field Tracker. Each item is who / what they ran / mechanism / new-or-repeat / link / tag.

Zero items would have been legal. This packet is not zero.

Items

  1. Matt Palmer (@mattyp) — restated Demo/Content Bot and the Instacart/Amazon pair; added an overnight Wii-baseball emulator grind. Mechanism: the cloud computer keeps working with the laptop closed; approval sits in front of the Cursor Cloud Agent write. Repeat of the 11 Aug essay, plus one persistence specimen that is not portable. https://x.com/mattyp/status/2087314806596641159 — deposit raw/sources/2026-08-11-mattyp-content-bot-overnight. repeat on Demo/Content/handoff; killed-usage on grocery (already refused); overnight grind is a toy, not a seat.

  2. Eric Zakariasson (@ericzakariasson) — 100 setups seen by launch day, three-post thread. Mechanism: a catalog is not a lane. Portable rows, already named 13 Aug: drive the GUI when the API permission is missing (#34, #41); reconstruct a promise from contracts + chat + CRM and flag the questions only a human can answer (#39–40); human gate before spend (#8, #75–78); quiet-when-nothing (#70). New-to-bank: #50, vet the fleet for spend and kill the wasteful ones — a Steward canary. Almost everything else is Salesforce / Gong / LinkedIn / Outreach. https://x.com/ericzakariasson/status/2087258914060664902 — deposit raw/sources/2026-08-11-zakariasson-100-use-cases. bank-candidate on #34, #39–40, #50, #70. repeat on the rest as a menu.

  3. Nate — twelve Bots in about eight hours. Rule he kept: a Bot owns a theme, not a task. Shared Linux box as the decision that makes the roster work and as the security boundary. Mom login handoff. Full essay paywalled (natesnewsletter.substack.com/p/grok-bot-review); claims above are from the public table of contents plus the pinned-list gloss, not from the body. bank-candidate on theme-not-task, with a watch: theme is wider than one duty and can become General Helper. repeat on shared-computer + login handoff.

  4. Ruben Hassid — built a daily AI-news digest, then asked it to send via Gmail. Mechanism: quiet digest of a scored beat, same shape as Brief. Method is Gmail on the shared computer, which this vault’s public-only line refuses. Caption in the post calls the VM “its own computer” — the FAQ says one computer per user. Public half of https://ruben.substack.com/p/what-the-f-is-an-agent (2026-08-12); paid section not retrieved. repeat on digest; killed-usage on Gmail-as-method; killed if anyone files “each bot has its own computer” from the screenshot caption.

  5. Official docs, re-fetched 2026-08-15 — get-started and FAQ. New citation: a first task names outcome, sources, constraints, deliverable, and review point. Confirmed, no change: one computer per user; do not use Bots as a security boundary; delete does not clear files or logins; Teach a task is ten minutes and gradual; Linux / Android / iPad unsupported; weekly usage with on-demand overage; Legacy Privacy Mode blocks the product. bank-candidate on the five-part first-task template. repeat on the rest.

  6. Atomic Bot explainer (Andrew Dyuzhov, 2026-08-13) — competitor page, useful only for the killed claim. Quotes the FAQ and the security line; notes one HN beta user still saying “separate VM per bot” against those pages. Secondary: Teach-a-task held on short recordings and was hard to debug when a ten-step routine broke; no model picker (teams docs, not independently opened this pass). https://atomicbot.ai/blog/what-is-grok-bot killed restated on per-bot computer. bank-candidate on short-vs-long Teach-a-task, tagged as second-hand until a named runner says it.

  7. Emily Gavrilenko / Ariv Gupta — LinkedIn. EA + specialists + human gate; named chief-of-staff Bot “sandeep”. Bodies not retrieved (auth wall). Filed as named, not quoted. repeat of fleet-plus-gate / Finn-adjacent CoS if the posts are what the titles say.

  8. jjcm (HN 49261514) / VentureBeat on Shumer — GUI-as-API over WhatsApp with a token-burn ceiling; Shumer researcher + writer + CoS. HN returned Sorry.; VentureBeat hit a Vercel checkpoint. Not quoted. unfetched.

Killed, still dead

  • Each Bot gets its own cloud computer. FAQ 2026-08-15: “The computer is assigned per user, not per Bot.”
  • Separate Bots are a security boundary. Same page, same sentence as 13 Aug.
  • Trust-agents-more-than-ourselves. No new source revived it.

What this packet does not do

Wiki ingest is still the 13 Aug map. This packet feeds the bank. It does not create Field in the product. Watch and Brief still meter first.

Grok 4.6 half of the list

Official model page, launch post, Copilot note, reasoning (xhigh), release notes, model card. $2/$6 unchanged. Consumer Grok surfaces still later. Does not reopen Grok 4.6 on the frontier, not the lead.

Sources

  • Pinned list
  • Palmer overnight
  • Zakariasson 100
  • Palmer intro (already compiled)
  • docs.x.ai/grok-bot get-started + FAQ, fetched 2026-08-15
  • Ruben Hassid, public half, 2026-08-12
  • atomicbot.ai/blog/what-is-grok-bot, 2026-08-13

Later packet (first-party playbooks, not a replacement): Grok Bot Field Packet 2026-08-31