Standing Research Agents
Standing Research Agents
The research here splits into a standing half and a session half. Four always-on cloud agents hold the standing half — watching, fetching, filing. Execution stays with the local agents on the Mac, and judgment — what a finding means, what becomes a page, what ships — stays at the desk.
The cloud agents see only the published repo — exactly what any stranger can clone.
The four lanes
The cloud side runs on Grok Bot, where every agent on an account shares one persistent computer. Four agents, one lane each.
Watch holds the estate’s standing checks: the public sites answering, deploys finishing green, search indexation moving, links staying alive. It speaks only on exceptions — a silent day means a healthy estate — and it never touches production. It reports; repair happens at the desk.
Brief reads X and the open web each morning on the beats this vault works — AI and agentic engineering, learning science — and files one exception-first brief: what changed, what crossed a threshold, then the roundup.
Intake feeds the research banks. It sweeps the standing sources — feeds, paper servers, the channels worth following — scores each find for relevance and novelty, and files queue deltas for review. Banks like the Two Egos Research Bank were gathered by parallel agents inside sessions; Intake runs the same collection as a standing lane, so the banks fill between sessions instead of during them.
Corpus audits the published wiki itself: near-duplicate pages, claims that contradict across pages, dead wikilinks, pages missing their sources. It files review packets and never edits — every change to the vault passes through a human hand.
One computer, one trust line
All four agents share one cloud computer, where files, logins, and credentials are common property and files outlive any single agent. That geometry sets the trust line: the cloud side carries only what is already public. A grocery cart, a mail session, a shopping login, or a spend that does not stop for a person is the usage that line exists to refuse. Other people’s write-ups of those setups are field evidence, not a roster to copy. A helper that sweeps a public feed and files a packet is the same shape as Watch and Brief; a helper that shops is the opposite shape. The vault’s private half — drafts, raw sources, finances — is gitignored and structurally out of the agents’ reach; they clone the published repo and see what a stranger sees. Execution runs closer to home: the local agents on the Mac hold the disk, the credentials, and the build stacks, with Claude Cowork carrying synthesis and structure and Grok Build carrying toolchain work. Every irreversible act — a page published, a deploy, a vault write — passes the desk.
This is the capability lens run as an org chart. The standing lanes are its Delegate cell held as duty, and session work is Augment — the pages themselves are its clearest case: agents draft them under the vault’s writing standards, the model’s default selling voice is rejected on sight, and final cut stays at the desk. A page is a position its author holds, whoever typed the first draft.
What the structure costs
One shared computer means one security domain: a credential handed to any agent is visible to all of them, so the agents hold fresh, scoped, read-only keys and never the Mac’s. The quota is weekly and metered past its included allowance, so every lane pays rent — a lane whose output stops changing what gets read or done is retired rather than left running. And the field’s record with unattended agents is poor: scheduled briefs measurably drift generic within a few weeks, and unmonitored agents rot silently while still reporting success. Every lane therefore carries a canary — a freshness check a stale run would fail — and a standing review date. The working test is concrete: deploy failures surface as alerts instead of by hand, the banks gain entries between sessions, and the briefs stay worth the two minutes they ask.
Sessions stay what they were — the place where findings become positions and positions become pages. What changed is the other half. The watching, fetching, and filing no longer wait for a laptop to open: the standing lanes run through the night, the review packets are there in the morning, and the research this vault wants conducted is being conducted whether or not anyone is at the desk.
Sources
- Grok Bot documentation — the shared-computer architecture, routines, and quota model: docs.x.ai/grok-bot. FAQ re-fetched 2026-08-31: one computer per user, not per Bot.
- Matt Palmer, Intro to Grok Bot, 2026-08-11 — first-party practitioner essay. Sweep-and-file specimens match Watch and Brief. Grocery and delivery specimens are the usage the public-only line refuses. Bank: Grok Bot Practitioner Bank.
- First-party playbooks on x.ai/bot/guides, captured 2026-08-31 — field evidence of a chief of staff and of mail, ads, and store logins this drawing refuses. Packet: Grok Bot Field Packet 2026-08-31
- On scheduled content drifting generic within weeks: Things I built with AI that completely fell apart
- On unmonitored agents degrading while reporting success: You can’t train an AI agent and then just go away, a taxonomy of silent agent breakage
Related
- Human vs AI Capability Lens
- Grok 4.6 and Grok Bot — the teammate against the model and against Grok Build
- Grok Bot Primer — the live setup this drawing became
- Bot Operating Rules — report-only, one-finder, finding-as-spec
- Grok Bot Field Packet 2026-08-31 — the maker’s how-to pages this drawing is choosing against
- Current Agentic LLM Stack
- Automation and the Job Iceberg
- The Two Meanings of Ego