AI marketing manager agent: Night Shift, on the Hermes runtime

The marketing manager agent, Night Shift, is an AI agent that does the work of a marketing manager for AR Labs, under a person's authority. An AI agent is a language model run in a loop with tools, so it can act as well as reply. Night Shift runs continuously on a server, keeps its own records, and works with Richard M. Thompson, who founded AR Labs, on a fixed division of work: Richard writes, and the agent researches, recommends, briefs, coordinates and measures. Nothing is published, sent or spent without his approval. This page describes how the agent is built and run; the lab's own marketing plans and findings are not published.

The division of work

The agent tells the person what kind of piece to write next, for which readers, why it serves the current goal and how it will reach them. The person supplies the writing and the voice. Drafting publishable prose is not the agent's default assignment.

The person	The agent
Agrees goals and decides	Proposes goals, approaches and the next useful action, with evidence
Writes the content	Prepares a writing brief for each recommended piece
Approves anything published, sent or spent	Prepares distribution and follow-up within that approval
Reviews recommendations	Researches markets, buyers, competitors and marketing method
Answers when a decision is needed	Keeps the plan and its next actions moving across sessions, and measures results

Each brief carries the current goal, the intended reader, the question or obstacle the piece addresses, a recommended format and angle, supporting evidence or examples, the next action the reader should take, how the piece will reach readers, and how and when its result will be assessed. Measurement follows the path from exposure among relevant readers to responses, conversations and engagements. Reach, rankings and clicks help diagnose a channel but do not show that it produces work.

The agent's standing responsibilities are to agree goals with the person, maintain the routines and channels that serve them, run campaigns under agreed approaches, measure and diagnose results, and turn repeated successful work into reusable procedures.

How it runs

Night Shift runs on Hermes, Nous Research's open-source agent runtime, as a background service on a rented Linux server. It has an identity file that states its role and what needs approval, a workspace that is its own git repository, and its skills: packaged instructions, with any scripts, that an agent loads only when a task needs them.

flowchart LR
  P["A person"]
  LS["Laptop sessions"]
  MB[("Shared mailbox")]
  subgraph SRV["Server, always on"]
    NS["Night Shift on Hermes"]
    SK["Marketing skills library"]
    WS["Workspace: working state, plan, evidence"]
    CP["Read-only copy of the lab's repository"]
  end
  TS["Task store"]
  WEB["Web search and scraping"]
  P <-->|"messages, approvals"| NS
  LS -->|"briefs"| MB
  MB <-->|"requests, replies"| NS
  SK --> NS
  CP --> NS
  WEB --> NS
  NS --> WS
  NS --> TS

Its marketing method comes from Corey Haines's public marketingskills library, which held 53 skills when we installed it in November 2026. We installed the whole library as its author intends rather than adapting skills one at a time, so its command-line tools and integration guides arrived with it, and we adapt only where a skill collides with the runtime. One skill keeps the shared marketing context, another turns research into a plan that is reviewed and can be resumed, the specialist skills carry out the work, and a routines skill covers recurring work once it is proven. A web search and scraping service, Bright Data, is connected through MCP, the Model Context Protocol. A copy of the lab's main repository, refreshed nightly and treated as read-only, gives the agent company goals and evidence of what the lab has built, so it can label a capability as described, measured or only planned.

How its work survives between sessions

A session of the agent ends, and the next one starts with nothing in its context window. Three kinds of record carry the work across:

• A working-state record in the workspace, the agent's resume point: what is under way, what was agreed and what comes next.
• A task store in Linear, an issue tracker, with issues and sub-issues for the work, "blocked by" relations for its order and comments for progress. We chose it after the agent's built-in to-do list proved to last only one session. The runtime's own kanban board was set aside because its parent links mean prerequisites rather than containment and it is built for unattended workers.
• Evidence folders holding research notes, each finding labelled by how it was obtained, and the plan documents the marketing skills produce.

The test for all three is that a fresh session can resume the plan at the right task, and derive the next writing briefs, from the records alone, without the person re-explaining.

How it receives requests

Requests reach the agent two ways. Sessions on the laptop write a dated file into the shared mailbox, and the agent replies with a file and leaves a receipt; the file is the record. Conversation happens in messaging: a direct chat with the person, and a thread in the lab's shared channel where another party mentions the agent by name to ask it to act. Vox Prime, the identity laptop sessions use in that channel, is on the agent's list of permitted senders. The agent addresses a role, such as the strategy agent, rather than whichever agent currently fills it.

Authority limits

• Nothing is published, sent or spent without the person's approval.
• A draft of the shared marketing context stays a draft until the person confirms it, and the agent's proposals stay open until the person accepts them. Facts, hypotheses and approved decisions are kept apart.
• The agent owns its instructions, skills and records. Changes to them arrive as briefs in its mailbox, which it installs itself.
• The agent cannot edit its own credentials or runtime configuration; the runtime's protections prevent it. An authorised operator maintains those, using the runtime's supported commands and keeping a private backup first.
• Between those limits it keeps working. It does not ask for confirmation between internal research and preparation steps; it stops when a decision, a missing contribution or an approval is required, says exactly what it needs, and carries on with work that does not depend on the answer.

How it was designed

The design started from the role. We drew the duties from published descriptions of a marketing manager, the O*NET occupational profile and the Chartered Institute of Marketing's professional framework, before deciding what the agent should hold.

The structure we designed for its work is the spine model, built on logical levels, Gregory Bateson's idea that each level of a system sets the frame for the level below. The outermost frame is the company: its goals and the interfaces through which marketing reaches the rest of the lab. Inside it is the marketing frame, holding the agent's role, mandate, authority and standards. Inside that runs the spine, a chain in which each item has exactly one parent: a marketing goal, an approach to it, and the campaigns and routines that carry the approach out. Tasks attach to whatever they serve. Stores sit beside the spine and outlive the work: evidence, decisions, lessons, a library of assessed resources, and procedures.

flowchart TD
  CC["Company frame: company goals and interfaces"]
  MF["Marketing frame: role, mandate, authority, standards"]
  MG["Marketing goal"]
  AP["Approach"]
  WK["Campaigns and routines"]
  TK["Tasks"]
  ST[("Stores: evidence, decisions, lessons, resources, procedures")]
  CC --> MF --> MG --> AP --> WK
  TK -.->|"attach to"| WK
  ST -.->|"cited by"| MG

Each link from a marketing goal to the company goal it serves, and from an approach to its goal, is a stated hypothesis with evidence for and against and a verdict: untested, supported, contradicted or inconclusive. Marketing goals are written as objectives and key results; routines carry their own measures. Evidence is labelled measured, provided by the person, calculated, estimated or proxy, and a later record never upgrades the label without new evidence. A lower level never edits a higher one: when work meets a contradiction with a level above, the agent raises it there as an item, with the evidence.

The design is accepted in fifteen recorded decisions and its implementation is deferred. We chose to start with the installed library and the existing working records, and to add structure only where use shows a need.

Keeping an agent like this on course

These practices come from running the agent on real work, and we now build them in.

• A finish line and a date on every assignment. A goal names its outcome, its measure and its time horizon, and an assignment lists the conditions under which it is done. An internal target date is a planning aid, not a sign that unattended execution is switched on.
• Name the method in every brief. When only part of the method library was installed, the agent filled the gap with the nearest method it had, a software planning skill. Installing the complete library, and naming the method for each piece of work, closed that gap.
• Check results against sources. A development session reads the agent's saved work in its live workspace to verify it, and records what that read did and did not repeat. Completing the sections of a template does not validate a finding, and a capability is labelled by the evidence for it.
• One home for each rule. Counting what the agent read before it said a word showed the same rule restated across several files in different words. Logical levels give each rule one place, so the agent can tell which rules apply.
• Durable records over session memory. Plans and tasks live in files and the task store, never only in a conversation.
• Instructions load at the start of a session. The runtime reads the identity file and workspace guide once, when a session begins, so every change to the agent's instructions ends with starting a new session.

What exists today

Running: the agent as a service, with the whole skills library installed, web research connected through MCP, a persistent task store, the mailbox, and a direct chat with the person. On 9 November 2026 a laptop session posting as Vox Prime sent the agent a request in the shared channel and read its reply, and the agent left a receipt in the mailbox. Its first supervised assignment, the groundwork a marketing plan needs, planned with the library's own methods, is under way.

Designed and not built: the spine model; scheduled work selection with execution budgets, so that repeated work proceeds without a person prompting each step; and a reporting route to the strategy agent, expected to be a dated file in the mailbox that a session routes into the operational state.

Planned: once the method has proven itself on the lab's own work, packaging it as an installable marketing manager that another owner of a Hermes agent can run for their own business, with the method kept separate from any one business's data.