From Dispatcher to Work Chief of Staff in a Multi-Agent System

Before routing ambiguous work, clarify the outcome, identify unknowns, choose the smallest structure, and define an explicit handoff.

# From Dispatcher to Work Chief of Staff in a Multi-Agent System

A vague piece of work should not be sent immediately to the “closest” specialist. Before routing, the system needs to help the person clarify the desired change, identify what remains unknown, and choose the smallest useful next step. This article calls the role responsible for that work the Work Chief of Staff: it clarifies, challenges, classifies, and recommends routing, but it does not make or execute decisions that require authorization.

Terminology note: Work Chief of Staff is not an official Hermes title. It describes a work-intake role, not an autonomous decision-maker; the example below demonstrates the method.

When a dispatcher is not enough

Dispatchers work well when a request is already clear: identify the appropriate specialist and send the work. Many requests begin differently: “I want to start a project, but I am not sure whether I have defined the problem correctly.” Routing too early forces a specialist to guess. Stopping there does not help the person make the work testable.

Hermes Kanban provides routing components: named Profiles, persistent work queues, and Tasks that can carry an assignee, parent, workspace, priority, Skill, runtime, and retry limits. Those components can hold work, but they do not decide what the real problem is. [Kanban documentation](https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/features/kanban.md)

The four-step work-intake model

  1. Clarify the outcome. What state should change, who needs it, and what counts as done?
  2. Identify the unknowns. Is the gap evidence, options, decision authority, resources, or execution method?
  3. Choose the smallest structure. A Task, an existing Board, a new Board and workspace, a proposed Skill, or a research question that should not be routed yet.
  4. Route explicitly. State the outcome, sources, constraints, responsible Profile, acceptance criteria, and reporting format.

This model does not make the entry point an all-purpose expert or an autonomous CEO. It makes uncertainty visible before work is assigned.

A fictional example: turn an idea into a testable task

Consider: “I want to research how remote work affects team collaboration, and maybe turn it into a report.” Before assigning it to a research Profile, the audience, decision to support, scope, and report format are still unknown. A smaller next step is a problem-definition Task: ask the research Profile to collect official statistics and research reviews, without deciding the conclusion for the user.

| Field | Initial content | | --- | --- | | Goal | Decide whether a type of team should adjust its remote-collaboration norms | | Known | The user wants to study collaboration and may produce a report | | Unknown | Audience, time range, method, and deliverable form | | Smallest next step | A problem-definition Task, not a new Profile | | Human checkpoint | Confirm audience and evidence threshold before expanding into a Board |

If source collection shows the topic deserves sustained work, create a Board and workspace. If a consistent screening method later emerges, consider a Skill. The appearance of a topic does not require a permanent new role.

A Session is evidence context, not knowledge or a correct decision

Sessions can preserve message history, model settings, tool calls, and time information. That helps answer why a classification was made, but an archived record does not make the decision correct or turn it into reusable knowledge. Important choices still need a Task record that exposes purpose, sources, unknowns, and human checkpoints. [Sessions documentation](https://github.com/NousResearch/hermes-agent/blob/main/website/docs/user-guide/sessions.md)

Limits and approval boundaries

A Work Chief of Staff should not create a Profile, start a Job or Cron, send something externally, or make another consequential choice without explicit human approval. It should not hide weak evidence behind tidy categories. Evaluate the model by tracking routing accuracy, operating cost, clarification outcomes, and human corrections. Start with small, reversible work before changing the role or workflow.