Skip to content
outstaffoutstaffOutstaff

0003: The office schedules all work

Status: accepted

Context

Section titled “Context”

There are two simpler ways to start agents, and both break down.

Let the issue tracker start them. Every agent wakes on every status change. Assignment then depends on what the tracker allows. Pull request events need CI jobs, and those jobs often can’t reach a self-hosted tracker.

Let agents react to events on their own. No single place knows the state of all the work, so loops and stalls go unnoticed.

Decision

Section titled “Decision”

Each team runs one office container.

Intake. The office takes in events from the code host, the tracker and chat:

  • by polling (the default, which works behind NAT);
  • or by webhooks, when the office is reachable.

Work state. It keeps current state for every ticket, pull request, CI run, review and stack, linked to the people involved.

Workflow engine. Rules decide what each event means (see 0011; since 0018 the event pipeline does).

Work items. Anything that needs a person becomes a work item: an assignment, a review round, a conflict, failing CI, a question from another agent, a chat request, or a scheduled job.

Scheduling. The office schedules work items per person:

  • human chat first, then human-triggered work, then agent-triggered work, then scheduled work;
  • a loop guard limits only runs triggered by agents or automation;
  • a usage policy defers scheduled work, then pauses automatic work, as LLM usage limits approach;
  • humans are always answered.

Agents connect out. Agents open an outbound connection to the office. They need no inbound port, and the office needs no access to their containers.

Messaging. Agents ask each other questions through the office (ost ask). The question becomes a child work item, and the answer resumes the asker’s work.

The tracker is optional. Without one, work items in the office are the record. With one, agents keep their tickets in sync.

Consequences

Section titled “Consequences”
  • No CI jobs or self-hosted runners are needed in client repositories.
  • The office needs its own identities: a read-only code-host app, a tracker account and an optional chat bot, each named <team>-outstaff-office.
  • The office is stateful (see 0007).