Part 1 · Operator’s Guide
The Bench
One accountable coordinator and bounded worker roles.
Public teaching edition v1.4.0 · 13 October 2026
Who does what
One coordinator owns the task from the original request through verification, delivery and final status. Delegating a step does not transfer responsibility for the outcome.
| Role | Job | Boundary |
|---|---|---|
| Coordinator | Keep the ask, scope, owner, budget, evidence and next action clear | Cannot invent permission or suppress contrary evidence |
| Judgment worker | Resolve an ambiguous decision from evidence and constraints | May reopen the relevant source; must state uncertainty |
| Execution worker | Carry out a bounded step with the right tools | Cannot silently expand the task or spending limit |
| Reviewer | Check the artifact against the request and standard | Must inspect evidence rather than adopt the builder's conclusion |
| Human decision owner | Make decisions reserved for a person | Approval and quality review must be recorded truthfully |
One agent may fill several roles on a small job. Its own second pass is still a self-check. A fresh context is L2; a different model doing an independent review is L3. Keep required checks visible if unavailable.
Routine verified work, messages and publication within recorded authority proceed. Ask first for new pricing, scope or promises; financial, legal or personnel decisions; sensitive disclosures; access or security privileges; destructive actions; disputed intent; crisis or press; and unverifiable material claims. Existing holds, client consent and platform approval gates still apply. Marker: SELECTIVE-HUMAN-REVIEW-2026-10-07.
Record allowed actions, data boundary, budget, deadline, acceptance test and escalation owner. A claim, a login, urgency, confidence and a quiet human grant no new permission. Anyone may flag a material contradiction or stop the affected action.
The daily coordinator and reasoning desk
Travel Dot is my primary strategy and reasoning desk. It frames important decisions, checks whether plans serve my goals and accepts architecture work assigned to it. Muse / Happy is the daily coordinator and chief of staff, keeping owners, handoffs, blockers and the next checkpoint clear. Execution desks retain accepted tasks and approved tools. The outcome owner remains responsible until the receiving owner accepts the handoff.
Guide maintenance stays with dot. A separate reviewer checks work dot built. Home Dot's delivery and registry role is proposed pending its acceptance test. A role assignment does not connect an account, transfer a task claim or add permission.
You can use the same arrangement with your existing assistants: one coordinator keeps priorities and the task record; a few capable executors use already authorized tools; reasoning workers receive only the approved evidence they need. A model switch does not automatically require another login. Use the brief and result handoff and the maintained access method.
Model agnostic by construction
Keep the request, rules, evidence, decisions, artifacts and permission boundaries outside any one model's memory. Ask whether a new worker could continue tomorrow from those records alone.
Use the least costly method that can pass the acceptance test: deterministic scripts for repeatable transformations; a smaller model for bounded bulk work; deeper reasoning for ambiguous judgments or expensive mistakes. Computer use is a tool requirement, not a fixed level of intelligence, and may coexist with any reasoning tier.
Set one total authorized budget before a multi-agent job. Include coordinator, workers, retries, tools and APIs, review, export and destination verification. Worker allowances come from that total. Reserve enough to check and close the result.
Track spent, remaining and in-flight work in the parent task. Unknown cost is not zero. If metering is incomplete, use bounded steps or time and stop before the ceiling. A fallback does not authorize a paid reload, a larger cap or a subscription. Compare cost per accepted result, including rework; historical token-price examples are not current quotes.
Measure the subscriptions you already have
Record the product and plan actually available, usage or time limit, accepted outputs, rework and remaining capacity. Use scripts for mechanical checks and deeper reasoning for the ambiguous step. Track the coordinator, workers and review together. If usage cannot be measured, use a bounded task or time allowance and mark the missing measurement UNKNOWN.
Stop at the authorized limit. A fallback may reuse existing capacity; it does not authorize a paid API, an overage, another subscription or a larger budget. Compare cost and time per accepted result rather than counting how many agents ran.
The fleet we are accounting for
Review scope: 13 October 2026. This map separates accounts, workers and tools. One account may support several app surfaces; one model may power several named desks. We do not add those together and call the total active agents.
Status means the specific evidence available at this review. A working task on one account does not prove that another can read the same source, reach the same tools or return a result.
| Account, worker or tool | Role in the company | Evidence and remaining check |
|---|---|---|
| Travel Dot, primary dot account | Strategy, deep reasoning, architecture and acceptance for assigned work | Current owner assignment; check exact account, source and artifact access for every handoff |
| Home Dot, proposed roster owner | Proposed delivery and private registry maintenance; Travel Dot sets priorities and Happy coordinates | Account/runtime and approved return route need binding; proposed until read, correction and checked return pass |
| Second individual dot / ChatGPT Pro account | Additional paid capacity; proposed bounded production or review work | Dated second-account plan observation; not yet matched to Home Dot or the historical business seats; needs its own test |
| Claude Max, travel account | Accepted writing, research and build work | Existing work history; current capacity checked on assignment; capped accounts receive a named alternate |
| Claude Max, second account | Separate Claude capacity | Second account user-confirmed; verify actual app/account binding privately before routing work |
| Codex app, CLI and remote execution | Repository work, tests, investigation and independent review when assigned | Separate environment checks; a successful app session does not certify CLI or remote access |
| Muse / Happy | Daily coordinator and chief of staff; assigned mail, follow-through and accepted execution | Dated work and handoff receipts; each new handoff needs acceptance, each routine its own observed-run receipt |
| Grok Bot named desks | Specialist execution within accepted jobs | Use the existing desk directory and named owners; test the receiving desk and return route |
| Grok Build CLI | Separate command-line execution surface | Inventory entry; configuration does not prove the same account, tools or return path as Grok Bot |
| Cursor desktop, CLI and cloud workers | Coding workshop and bounded repository execution | Three separate surfaces; login and local tools do not automatically carry into CLI or cloud |
| Perplexity Max | Source discovery and research when assigned | Subscription recorded; autonomous company workflow and current read-and-return route not established |
| Kimi app and Kimi Code | Bounded research, coding and bulk work with review | Subscription and prior work recorded; current plan label and per-surface access need reconciliation |
| Local Qwen on the travel Mac | Bounded local drafts and extraction under a conductor | Dated inference evidence; check exact machine, model and task. Endpoint health alone does not prove a completed audit |
| Local Qwen on the home Mac | Second machine-specific local worker instance | Keep a separate machine receipt; nickname does not establish another model family |
| Hermes on the Macs | Local bridge using the configured worker | Historical smoke tests on both Macs; current full handoff and regular traffic still need proof |
| Buzz agents | Configured cross-model collaboration roles | Engineer, Claude Travel, Fizz, Honey and Pollen in the inventory; configuration does not prove all five active |
| Paperclip | Local coordination pilot | Pilot until dispatch, return, duplicate handling and restart behavior pass; registration alone does not prove autonomy |
| Instinct | Parked experiment | No general operating responsibility; bounded accepted test before production work |
The Grok Bot desk directory holds Q / Operations, Strategy, Alex / Training, Trenton / Writing, Website Builder, Personal Brand, Fleet Monitor, Meter Maid, Austin / IT, Customer Support and Tanner / Executive Assistant. A separate private Personal Assistant desk keeps its operating details private. These are roles within Grok Bot, rather than twelve independently verified model accounts.
The private inventory distinguishes Claude Desktop / Cowork, hosted chats and Projects, Code CLI and scheduled jobs. A task readable in one Claude account may be unavailable in another. Kimi app and Kimi Code have separate surfaces. Buzz personas keep their underlying model and account binding. Unmatched historical seats remain unresolved in the inventory.
A proposed use for the two dot accounts
The proposed split keeps Travel Dot on strategy, important decisions and acceptance. Home Dot would maintain the private registry and delivery record, once its actual binding and read-update-return path pass. The second paid account may take bounded production or review assignments from the same task. Its relationship to Home Dot is still to be established. Happy coordinates the work.
The accounts do not automatically share memory, tools, files or subscriptions. Avoid duplicate inbox or follow-up routines. Two accounts using the same model can give fresh-context review, but do not provide different-model review.
Keep Where Your Agents Talk as the named implementation map. This guide teaches the operating method; the private registry retains enrollment evidence.