Part 2 · Deep Reference

The Organization and Its Public Face

Who holds the work, who answers for the result, which choices stay with people, and where to read next.

Public teaching edition v1.4.2 · 15 October 2026

AI agents can do a lot of the work in a business. But every result still needs one owner, and some choices still belong to a person. If nobody owns a result, the founder ends up owning all of them. This page shows who holds each job. It shows how to ask a person for a decision in one short message. It also shows how to keep each client's private details safe. A dated reading list comes at the end.

Who holds what

  1. Agentshold tasksClaim one task, do the work, and attach proof of where it landed.
  2. An accountable functionowns the outcomeOne owner per result, such as publishing, client reporting or QA. The owner is a function, not a favorite model.
  3. A named persondecides reserved mattersFor example: new pricing, scope or promises; money, legal or people decisions; access or security; destructive steps; crisis or press. Full list.

Handoff rule: the coordinator keeps the task until the receiving owner accepts the same record and can continue with the evidence, tools and authority it needs.

Work moves through one task record. When a decision is needed, pass along only the decision, one level at a time.

Humans and clients

A team needs a human entry path and an agent entry path into the same authoritative method. Both should identify the owner, working record, boundaries and acceptance criteria. For example, people start at New Team Member — Start Here and agents start at New Agents Start Here: same company, two front doors.

Use one accountable function for each outcome and a named human for its reserved decisions. A bounded escalation contains the actual decision, deadline evidence, options, costs or consequences, and the worker's recommended next step. It does not ask the requester to become the relay between agents.

A bounded escalation, ready to copy

A good escalation is small enough to answer in one reply. The A-player standard for project leads gives this example:

“Approve up to $300 for this tested campaign by Thursday; the live dashboard and rollback plan are here.”

A poor escalation asks “What should we do?” and attaches a pile of unsorted context. Use this shape instead:

Decision needed: one sentence naming the choice itself
Why a person: the reserved category, for example price, scope or a promise; money; legal; people; access; a destructive step; or a crisis
Answer by: date and time, and where that deadline comes from
If no answer: what stays paused, and which independent work continues
Option A: what it is, and its cost or consequence
Option B: what it is, and its cost or consequence
Recommendation: A or B, with the reason in one line
Evidence: links the person can open without asking anyone
After deciding: the function that carries out the decision

Here is the same form filled in. The business and the numbers are made up, so you can see the finished shape:

Decision needed: spend $300 on a two-week test ad for spring window cleaning
Why a person: money
Answer by: Thursday at 5 pm, because the owner wants the ad running on Monday
If no answer: the ad stays off; review requests and free posts keep going
Option A: spend $300, capped, and pause it any day. Expected: 10 to 15 quote requests
Option B: spend $0. Spring work comes only from repeat customers and reviews
Recommendation: A, because last spring a $150 test brought 9 quote requests
Evidence: the ad preview, the $300 cap setting and last spring's results, linked here
After deciding: the ads function starts or cancels the test and posts the receipt on this task

Silence never grants a reserved decision. Name what stays paused, and keep the rest of the work moving.

Client-visible communication belongs in the agreed channel, with sufficient context and accessible evidence. Their agents receive only their own authorized material and public methods. Keep internal diagnostics, another client's details and private relationship notes out of that projection.

Every status reply, to a client or a teammate, says which of four states the work is in. The Always Reply rule names them. Answer on the thread where you were asked, within 24 hours. If you cannot, say who covers for you. Another thread is not a reply.

State The reply includes
Done What shipped, where it lives and the proof
Doing The next date, on the same thread
Blocked Who or what is needed to continue
Parked Closed or not yours, with the living owner named

Two kinds of “parked.” These four words label a reply. A task record has its own states, and one of them is also called parked. On a task, parked means a dated wait with a reason and a recheck date, as the job packet explains. On a reply, parked means this thread is closed and a named owner now holds the work. Say which one you mean.

Service expectations must be human-confirmed. A reminder, standing brief or generated handbook cannot create a new contractual commitment. Label impact as directly observed, influenced or unknown when the evidence supports those distinctions:

Accountability at scale

Agents hold tasks; accountable functions own outcomes; named humans decide matters reserved for people. A coordinator stays responsible until a receiving owner accepts the same record and can continue with the necessary evidence, tools and authority.

Keep one queue with views over the same IDs. Derive the board row, note skeleton and receipt from the packet when possible. Do not create multiple follow-up clocks for the same unresolved obligation. The agent job board shows one working example: a single waiting room for work that has no owner yet.

Measures that mean what they say

Useful measures include tasks opened, accepted completions, known collisions detected, external records refused, required-review coverage, verification coverage, rework, cost per accepted result and time to close by class. Define each measure and its observable scope. A count of rejected records is not a count of sends prevented.

Measure What it counts What it does not prove
Tasks opened New task records in the one queue Demand that never became a task
Accepted completions Results the receiver or reviewer accepted, with proof attached Work marked done without proof
Known collisions detected Two workers caught holding the same task Collisions nobody noticed
External records refused Outbound items a gate stopped for missing proof or approval Sends prevented; the same message may have left another way
Required-review coverage Share of results that needed a review and received one That the review was good
Verification coverage Share of completions checked where they landed That the result helped the business
Rework Accepted results that were reopened or redone Why it happened; read the run note
Cost per accepted result All spend, including people, subscriptions, tools, retries and review, divided by accepted results Value created; unknown cost is not zero
Time to close by class Median time from claim to acceptance for one kind of task A fair comparison across different kinds of task

Test the client wall before granting access

Before giving a client's agent access, test the actual wall and its data projection. Public or client views must omit unrelated clients, private judgments and secret material. Model changes must preserve these boundaries.

  1. Build the exact view the client's agent will receive.
  2. Search it for other clients' names, private judgments, internal diagnostics and secrets. Any hit fails the test.
  3. Ask for something that must be refused, such as another client's file, and confirm the refusal.
  4. Keep the result with the access record: who tested, when, and what was checked.
  5. Run the test again after any model, tool or permission change.

Use the current named coordination roles and one morning snapshot. Other views link these procedures and the same task IDs rather than starting another board or clock.

Public reading map

These public sources provide background to the method. Read their dates and distinguish historical examples from the current task's instructions. This guide's templates are self-contained; private operational sources are not required for a beginner exercise.

Every link below was opened on 15 October 2026. The date shown is the last-updated date each page reports in its metadata at that check.

Question Public source Updated
Where do people start, and where do agents start? New Team Member — Start Here · New Agents Start Here 23 Sep 2026 · 24 Sep 2026
How do I answer so nobody has to chase me? Always Reply 17 Aug 2026
Who owns a result when several people help? A-player standard and RACI 24 Sep 2026
Where does work wait when nobody owns it yet? The agent job board 2 Oct 2026
What happens when someone does not reply? An unanswered ask never stops the work 6 Sep 2026
Where is the portable skills library? Public skills repository Changes often; check its commit history
How do internal receipts identify the worker? Agent disclosure 7 Sep 2026
What evidence supports done? The checker and definition of done 6 Sep 2026
How should existing authorized access be checked? Reusing browser logins 14 Oct 2026
Where is the one public page for each concept? Canonical directory 19 Aug 2026

Put it into practice

Choose the example that fits your next useful task. Use this guide for the current method and the articles below for focused explanations and dated experience.

These sources may evolve independently. A linked example does not install software, connect an account or grant authority to act.

Page history
  • v1.4.2 · 15 October 2026: the 24-hour reply rule from its source, a note on the two meanings of parked, a filled-in escalation and re-checked reading-map dates. The first two came from a review of the whole guide.
  • v1.4.0 · 13 October 2026: a section that ties the work to the Second Curve, and a refreshed date for the roles-and-channels page.
  • v1.1.3 · 8 October 2026: plain-English opening, a who-holds-what diagram, a copyable escalation, the four reply states, impact labels, a measures table, a client wall test and a dated, expanded reading map.
  • v1.1.1 · 2 October 2026: six focused practice links. The guide was first published on 1 October 2026.

Tie the work to the Second Curve

The direction comes first. For each approved goal, keep a short link from the outcome to work that can change it.

PartWhat to write
GoalApproved result and source of the decision
EvidenceCurrent starting point, date and what remains unknown
Accountable functionOne function that owns the outcome
Next useful taskSmallest accepted result that moves the goal forward
Worker and reviewerWorker with current access and the separate checker
Success measureObservable result and what we will inspect
Stop or decisionWhat stays paused when evidence, permission or agreed limit is missing

Use this chain for client results, authority and content, books and teaching, and the time and relationships I choose to protect. The private plan holds private priorities and relationship context. This public guide teaches the method.

Measure accepted results and their effect on the goal. A post published, task closed or busy model is activity evidence. Business impact needs its own source. Mark impact observed, influenced or unknown; keep proposed targets separate from actual results.

The morning view shows the few decisions only I need to make, agent-owned work and material blockers. Each row links to the same task and checkpoint. The team should be able to finish without making me carry messages between tools.