A Field Tracker

Verdict: how people are actually using agents gets its own standing lane — a Field tracker, one duty, packet only, public posts only — and it feeds a use-case bank the way Intake feeds the research banks. I am not folding this into Brief or Intake. Those two already have jobs. This one watches named setups, what broke, and what they kept.

The seed is already on disk. This morning I compiled Palmer’s Grok Bot essay into the Practitioner Bank — four lanes, confidence tags, the grocery pattern filed as refused usage rather than inspiration. That bank is one essay plus the docs. Two days of other people talking are sitting in posts, and they are the interesting part. I treated them as fleet threats. They are field evidence.

What they are actually saying

Palmer, who works on the product, is not selling a roster. He is selling a feeling he named as a trap first: chasing notes apps and custom agents until the construction ate the work. What he wanted was an agent that does the manual things that already live on a computer, without a server to provision. The specimen he likes best is Demo Bot. Every day it walks his X bookmarks, finds a piece of new technology, drafts a prompt in his writing skill, and waits. If he approves, a Cursor Cloud agent stands up a prototype in a tech-demos repo. Fifteen minutes later he binds a port and plays with it. “The point of this bot is to remove the friction from trying new things.” That is not inbox triage and it is not a morning digest. It is a bet that exploration dies on setup cost, and that an approval gate in front of a write is enough trust.

JoePro is running the opposite temperature and landing in the same geometry I already drew for Brief. Connect the tools you live in. One weekday morning digest beats ten chats. Event watchers that ping only when something needs you. Specialists, not one mega-bot. Show a workflow once, save it as a routine, do not build a factory. “Start narrow. Signal over completeness. Stay quiet when there’s nothing.” I almost discarded this as “we already have that.” We have the design. We do not have his two-week report on whether the silence held.

Will Baumann was in the private alpha. His favorite use so far is a PM agent that finds tickets with active engagement and no owner, then drives them to a close or onto the roadmap. Attention without accountability. That is the quiet-breakage problem in someone else’s tracker, and it is the first use case I have read this week that is not “the bot logs into my mail.”

Mike Gonzalez has been on it a couple of weeks and says he may retire his Mac mini. Two hundred and twenty-two PRs on a side project, almost all from his phone, the agent computer unblocked from iOS. I cannot verify the count. I can take the claim seriously as a claim: the laptop is no longer the execution surface, and the thing that made that possible is a machine that does not die when the chat ends.

Celso Pinto tried it and the clever part failed in a useful way. “A chief of staff bot doesn’t automatically delegate to the sub-bots. You need to tell it to.” Continuous chat, no task object, hard to hold a mental model of how many plates are spinning. That is the no-orchestrator ruling with a product fact under it, not a taste argument.

Someone writing as yrzhe spent a day fighting three bots that each handed back a /workspace path that does not open on his Mac, and could not tell whether they were alive or stuck until he typed “keep going.” Then he said the part that matters: the computer that does not shut off is the actual win. Yesterday’s work was still on the box this morning. He could SSH in. That is the architectural sentence the launch posts keep burying under “teammate.”

Eric Zakariasson posted a hundred use cases the day of the beta. Almost all of them are Salesforce, Gong, LinkedIn, Outreach — not my life. Two mechanisms in that pile are portable. One: drive the GUI when the API permission does not exist. Two: reconstruct a promise from contracts and chat and the CRM, and flag the questions only a human can answer. The rest is a catalog, and a catalog is not a lane.

Alex Finn’s five-step setup — dump your life into one agent, ask it to design the fleet, implement the routines — is the thing I will not do, and it is also a thing worth watching. The research question is what those rosters look like two weeks later. General Helper with five routines, or one-duty specialists. That is a competitor-diff item, not a setup step.

Why this is not Brief, and not Intake

Brief is the morning exception signal: what changed on the beats, silence when nothing moved. Intake scores pinned sources for three research beats — AI, learning science, wiki craft — and files bank-candidates. A field tracker does neither. It asks: what did a named person actually run, on what trigger, with what human gate, and did they still have it a week later?

Those are different obligations. Brief that starts collecting use cases will stop being exception-first. Intake that starts ethnography will stop being a scored source queue. I have watched that collapse before. One bot, two products, no canary.

The honest case against a new seat is still the stronger-looking one on paper. Structure A is not ruled. Watch and Brief are supposed to meter a week before anything else is born. New bots come from rent reports, not from a good afternoon. Brief already reads X. Intake already tags agentic. A third bot watching the same website is how a weekly quota dies, and Finn’s step two is the failure mode of this impulse wearing research clothes.

I am taking the seat anyway, as a specialist, because the product is different and the seed will rot. The Practitioner Bank will be stale in a week if nobody is assigned to watch the field. Session research is how I built the self-talk and ego banks, and it is the wrong cadence for a launch week that is producing new setups every morning. The cost is quota and a third X-facing routine. Pay it only inside the included allowance, and only after one manual packet proves a source list — Palmer, JoePro, Zakariasson, Gonzalez, Pinto, Baumann, and whoever Brief flags as a named setup rather than a vibe.

What the bot is

Name: Field. Working label. Not Tracker, which sounds like Watch.

Single duty: how people are actually using agents. Named setups, routines, human gates, failures, what they kept. Public posts and public essays only. Packet only. Never writes the vault. Never signs into anyone’s tools.

Not: a morning digest (Brief). Not a paper-and-source feeder (Intake). Not a fleet designer. Not an inbox cleaner. Not Yuedu.

Output: one packet, twice a week is enough — use cases do not need daily spam. Each item: who, what they ran, the mechanism in one sentence, whether it is new or a repeat, link. Tag bank-candidate when the mechanism is portable (persistent computer, quiet-when-nothing, ownerless work, teach-once, GUI-as-API, human gate before spend or send). Tag killed when a claim dies on contact (each bot gets its own computer; trust-agents-more-than-ourselves). Zero items is legal.

Canary: if three consecutive packets do not change what I open or what the Practitioner Bank absorbs, retire the bot. Retirement is the system working.

Clearance: the same as the rest of the cloud side. Public material. No Gmail, no Slack, no Amazon. Other people’s use of those tools is the research subject, not the method.

The people stay in the bank as people. A wiki page that later absorbs a mechanism can put the names in Sources. The lane itself has to keep who said it, because the whole point of field research is that a setup without a person is a feature list.

What would flip me back to “just tag it on Intake”: ten Field packets that are indistinguishable from Intake’s agentic rows — title, score, link, no setup, no gate, no failure. If the ethnography does not show up in the packet, the seat was a rename and it should die.

I have not created it in the product. This entry is the assignment. One paste, one pinned source list, one manual packet, then a Tue/Fri routine if the packet was worth opening. Watch and Brief still meter first if the weekly allowance is tight; Field waits a week rather than eating their canary. The ideas are good. The logins are not the ideas.

Status, 2026-08-15

Gold file and first packet exist. Pinned list: raw/sources/2026-08-15-grok-guides. Packet: Grok Bot Field Packet 2026-08-15. Bank lane field-2026-08-15 on the Practitioner Bank. Field is still not a product bot. Tue/Fri routine waits on whether this packet was worth opening.

Status, 2026-09-01

Second packet exists. Hub: x.ai/bot/guides. Packet: Grok Bot Field Packet 2026-08-31. Bank lane field-2026-08-31. Wiki absorb written onto Primer, Bot Operating Rules, Standing Research Agents, Fleet Structures, Grok 4.6 and Grok Bot, Current Agentic LLM Stack. Field is still not a product bot. Watch and Brief still meter first.