Grok Bot Fleet Structures
Grok Bot Fleet Structures
Candidate structures for the Grok Bot fleet — the always-on agents of Standing Research Agents, drawn here at configuration resolution. Nothing on this page is ruled. Structures get added as they are derived, and the one that survives contact with configuration becomes the record.
Structure A — five bots, single duties
One bot per duty, no personas, no orchestrator. The names are working labels.
Steward holds no lane of its own; it reads the other four bots’ run records.
Watch — the estate monitor. Hourly reachability on the public sites. On every push, a deploy check — and on failure, the build reproduced on the bot’s own box with the offending file named. Daily Search Console diff: canary crawl times, coverage flips, indexed counts moving off zero. Weekly link-rot crawl. Exception-only voice — a silent day means a healthy estate. Needs one fresh read-only Search Console key and nothing else.
Brief — the morning signal. Weekdays 07:00: one sweep of X and the open web on the vault’s beats — AI, agentic engineering, learning science — plus whatever Watch flagged overnight, delivered as a single exception-first push. What changed comes first, the roundup last, silence where nothing moved. The daily digest reborn on the platform whose absence killed it. Needs only its schedule.
Intake — the research feeder. Daily sweep of the standing sources — feeds, paper servers, followed channels — each find scored for relevance and novelty, filed as one review packet. Weekly, the competitor bank re-run and diffed. The packet is the product; nothing enters the vault except through the desk. Needs chat delivery to start, a git bridge later if packets should land as files.
Corpus — the wiki auditor. Weekly clone of the published repo, then the sweep: near-duplicate pages, contradicting claims, dead wikilinks, sourceless pages — one packet, never an edit. Standing sub-duty, the mold check: tracked instruction and config files grepped for resurrections of retired tools, every new spore flagged with file and line. Needs git confirmed present on the box; the repo is public, so no credential at all.
Steward — the quartermaster, created last. Weekly: quota burn against the included allowance, routine health, review dates coming due, and one line per lane on whether it earns its rent. A catalog of setups other people run includes a row: look at the whole fleet’s spend and kill the wasteful ones. That row is this report. If the owner is away long enough that the product asks whether routines should keep running, and nobody answers, those routines pause; Steward reports the pause. Administration, never orchestration — the proven remainder of the chief-of-staff idea. The product’s ceiling is fifty helpers and group chats combined, and fifty routines on one helper. Those numbers are not this drawing’s roster limit. The weekly meter is.
Shared clearance. All five sit on one computer where files and logins are common property, so all five hold the same trust level: public material only, scoped read-only keys only, and every output lands as a push, a chat message, or a review packet. Nothing writes to the vault, nothing touches production.
Rollout. Week one, Watch and Brief alone — the metering week, watching how fast two lanes eat the weekly allowance. Week two, Intake and Corpus if the meter allows. Steward once three lanes run. Before any of it, five minutes of in-product verification: one computer or one per bot, git present or absent, disk quota, the actual weekly allowance, push reaching the phone.
The case against A. Five bots may be two too many for a shared weekly quota — the meter decides, and the structure shrinks gracefully (Steward’s duties fold into Watch, Intake’s weekly diff into Brief). Single-duty bots also multiply routines to maintain: five bots at even three routines each is fifteen schedules carrying canaries and review dates.
Sketches not yet derived
- B — three generalists. Ops, Research, Admin. Fewer routines to maintain, muddier accountability per finding; field reports so far favor narrow bots over broad ones.
- C — two bots plus routines. One monitor, one reader; the audit becomes a monthly routine instead of a standing bot. Cheapest on quota, slowest to notice drift.
- D — first-party playbooks. A chief of staff in the middle, specialists around it, helper-to-helper handoff, mail and ads and store listings on the shared computer. That is what the maker’s published how-to pages run. It is a candidate only as a record of what this drawing is not. The logins those seats need stay off this account. Packet: Grok Bot Field Packet 2026-08-31.
Related
- Grok 4.6 and Grok Bot — the teammate these seats run on, against the model that shares its name
- Standing Research Agents — the same structure one level up, and the laws the fleet answers to
- Current Agentic LLM Stack — where the fleet sits among the other agents
- Human vs AI Capability Lens — why every lane ends at the desk
- Grok Bot Primer — the live setup; Structure A was never signed, the Primer is what runs
- Grok Bot Field Packet 2026-08-31 — first-party CoS+specialists filed as sketch D, not a roster
Sources
- Create and manage Bots — account cap of fifty bots and group chats combined.
- Skills, routines, and automations — fifty routines per bot; routines may pause after a long period away.
- FAQ — one computer per user. Re-fetched 2026-08-31.
- Named-runner catalog row on fleet spend: Grok Bot Field Packet 2026-08-15
- First-party playbooks as sketch D: Grok Bot Field Packet 2026-08-31