Appearance
Always-on DutyGraph Agent
The always-on DutyGraph Agent is a proposed, optional feature. Today Business Autopilot runs only when an advisor clicks Run a pass. With this feature, the same pass runs on its own: it wakes on a schedule or when something changes, checks the company's map for gaps, and drafts a few questions. The advisor still approves each question before it goes out.
WARNING
Planned. This page describes a design. Parts it builds on exist: Business Autopilot's pass and its caps, the background job that runs every five minutes, the notification outbox and the advisor's approval. The company event inbox, the agent memory and the scheduled pass are not built. The Status box at the end says which part is which.
Optional, off by default
The feature is a company setting. A company turns it on when the advisor and the sponsor agree. With it off, Autopilot works as it does today: the advisor opens the Readout and clicks Run a pass.
| Today: Autopilot on demand | Always-on (optional) | |
|---|---|---|
| Starts | The advisor clicks Run a pass | A daily schedule, or a change at the company |
| Runs where | In the advisor's browser | On the DutyGraph server |
| Remembers | Gaps already asked, kept with the company's Onboarding record | Gaps asked, answers, and the advisor's corrections, kept on the server |
| Sends | Questions the advisor sends | Drafted questions wait for the advisor's approval, then go out |
| Limits | 3 Autopilot questions a person a day, none if 3 are open, never the same gap twice | The same limits, checked on the server |
How a run works
The agent does not run all the time. It sleeps, wakes for a short run, and sleeps again.
mermaid
flowchart LR
E["Event or schedule<br/>answer came in, file changed,<br/>question approved, daily tick"] --> I["Company inbox<br/>one list per company"]
I --> R["Short run<br/>read memory, check the map<br/>with the Readout rules"]
R --> M["Update memory<br/>what was asked,<br/>what was answered"]
M --> O{"Result"}
O -- "Nothing to ask" --> Q["Stay quiet"]
O -- "Gaps found" --> A["Draft questions<br/>for the advisor to approve"]
A --> S["Send<br/>to the person's client app Home"]
classDef planned fill:#eeeeee,stroke:#888,color:#333,stroke-dasharray: 4 3;
class E,I,R,M,O,Q,A,S planned;Step 1
Something happens
An event goes into the company's inbox. Events are: a person answered a question, a Company Repo file changed, a transcript was added, the advisor approved a question, or the daily schedule came due. Each event is saved once, even if it arrives twice.
Step 2
One short run
One run at a time per company reads the waiting events together. It loads the agent memory and checks the map with the same plain rules the Readout uses: a goal with no number or owner, work held by one person, a manager with more than eight reports, a role profile with missing answers.
Step 3
Update the memory
The run records what it found, what it asked before and what was answered, so it never asks the same gap twice.
Step 4
Stay quiet or ask
Most runs end with nothing to send. When a run finds gaps, it drafts one short question per gap for the person who can answer it. The drafts go to the advisor's Inbox. The advisor approves, edits or drops each one. The approval goes back into the inbox and resumes the run, which sends the question to the person's client app Home.
What the agent remembers
The Company Repo is already the record of the business. The agent stores only what the Repo does not hold.
| Memory | What it holds | Loaded |
|---|---|---|
| Standing instructions | What the advisor wants the agent to focus on or avoid for this company | Every run |
| Asked log | Each gap asked, to whom, when, and whether it was answered | Every run |
| Advisor corrections | Notes such as "do not ask the CEO about pricing" | Every run |
| Run notes | What each run found and why it stayed quiet or asked | When needed |
The DutyGraph Agent panel stays on demand. It reads the same memory, so what a scheduled run found shows up when the advisor opens the panel.
Guardrails
- Start quiet. The agent earns attention slowly. A run that finds nothing new sends nothing.
- Three a day. At most 3 Autopilot questions a person a day. None if the person already has 3 open from anyone. Never the same gap twice.
- Advisor first. Every question waits for the advisor's approval.
- Proposes only. The agent never changes a Company Repo file, a role profile or a Task Card. Answers are filed the way they are today.
- No ranking. The agent does not track response times or compare people.
- The company's own AI key. If a later version uses AI to read files, it uses the company's approved AI connection, as the DutyGraph Agent panel does today.
Rollout order
- Company inbox. One event list per company, one run at a time, duplicates dropped.
- Scheduled pass on the server. The daily schedule adds an event. The pass, its memory and the 3-a-day limits run on the server.
- Approval resumes the run. An approved question is an inbox event. The DutyGraph Agent panel reads the agent memory.
- AI reading of files. A separate step in the Business Autopilot plan: AI reads Company Repo files and writes the questions. Plain rules stay the first check.
INFO
Status. Built (Partial): Business Autopilot's Run a pass on the advisor's Readout, with plain rules, the 3-a-day and "never twice" limits, sending to the client app Home and filing answers in the role profile; a background job every five minutes that runs Living DutyGraph items and other queues; the advisor notification outbox; the DutyGraph Agent panel, which only proposes. Planned: the company setting, the company event inbox, scheduled passes on the server, the server-side agent memory, approval that resumes a run, the panel reading that memory, and AI reading of files.
Source: gitbook/product/roadmap/always-on-agent.md in the DutyGraph repository.