The terms, the mechanics, and why AOS is shaped this way.

An archetype is a buyer-facing marketing engagement pattern. Each archetype has one workflow: the end-to-end process that delivers the outcome. There are eight archetypes (Acquisition, Retention / Loyalty, Brand Building, Media Optimization, Content Production & Distribution, Campaign Execution, Event Activation, Measurement & Analytics). Archetypes organize AOS around what the buyer asked for, not by competence domain.

A workflow is the end-to-end process for an archetype. One workflow per archetype. This is the SKU: what clients buy. A workflow is a defined sequence of tasks that takes inputs and produces an output. Each task is performed by an agent. The Operating Layer orchestrates the workflow; the Intelligence Layer informs each task.

A task is the atomic unit of work. Each task is performed by one agent in one execution. It has a defined input and a defined output, and it runs to completion once when invoked. Tasks are reusable across workflows; "Audience Segmentation" is a task that appears in multiple archetype workflows. Tasks do not loop; the repetition pattern is at the workflow level.

An agent is an AI worker that performs a specific task inside a workflow. Agents carry domain expertise for their task type. The same agent can serve multiple archetypes: the audience segmentation agent performs the segmentation task in Acquisition, Retention, and Brand Building workflows. Agents are governed by the Operating Layer's Agent Governance function and evaluated by Agent Evaluation.

An output is the deliverable end-state of workflow execution. The artifact or operating condition the client receives, counts, and measures. Examples: a brand book, a GMMM model, a campaign in market, a persona document, a monthly performance report. Workflows produce outputs; outputs do not decompose into smaller outputs the way workflows decompose into tasks.

The Context Layer is the set of parameters that shape how a workflow runs for a specific client engagement. Four parameters: B2B/B2C, Industry, Geography, Company Size. B2C acquisition and B2B acquisition are the same archetype with different context. The Operating Layer's Intelligence Assembly draws different intelligence for different contexts.

The AOS Operating Layer is the shared engine layer every workflow runs on, regardless of which archetype it serves. It is how AOS runs: machinery, generic across every workflow. Three core functions: Intelligence Assembly (draws the right knowledge into each task; governs agents; evaluates output), Gate Management (gates progress at decision points; binds the Worker; records provenance), and Cross-Workflow Coordination (routes outputs between workflows; schedules runs; validates learning).

The AOS Intelligence Layer is the body of knowledge AOS draws on at every workflow gate. The Intelligence Layer is what compounds; the un-clone-able asset that makes "smarter every time it runs" structurally real. The Operating Layer is constant; the Intelligence Layer grows with every workflow that runs.

(1) Functional intelligence: how to do marketing work for each task type; domain expertise codified per agent domain; cross-client. (2) Client intelligence: what only the specific client knows about itself (brand, voice, audiences, products, history, competitive landscape, business rules); per-client; loaded and maintained by the Worker. (3) Deloitte intelligence: firm-wide IP, proprietary methodologies, benchmarks, cross-engagement patterns; cross-client. (4) Learned intelligence: accumulated from prior workflow runs, validated through the Learning Validation Pipeline; some per-client, some cross-client; compounds over time.

The Operating Layer is machinery; the Intelligence Layer is knowledge. The Operating Layer's Intelligence Assembly function draws the right mix of intelligence from the Intelligence Layer into each task at the right time. The Operating Layer is constant across all engagements. The Intelligence Layer grows with every workflow that runs. Together they form the engine that powers everything underneath the work layers.

The buyer names an objective (acquire customers, optimize media spend, build a brand). AOS maps the objective to the right archetype or archetypes. Each archetype's workflow activates; agents perform tasks inside the workflow. The Operating Layer orchestrates; the Intelligence Layer informs. The client receives the output. The Learning Validation Pipeline captures what worked, and the system is smarter for next time.

The Worker is the human who interacts with AOS at the boundary of every workflow. Two phases. In Phase 1 (Configuration), the Worker provides context: assembles client intelligence, configures workflows, reviews and approves intake-agent output. In Phase 2 (Run), the Worker plays a subset of four roles depending on the workflow: Authorize (green-light a run), Review (binary go or no-go at a gate), Redirect (change direction at a gate), and Judge (taste and quality call). The Worker is the commissioner and the judge, never the co-worker.

A gate is a defined point in a workflow where progress pauses and requires evidence to pass. The Operating Layer's Gate Management function controls gates. At a gate, the Worker steps in to review, redirect, or judge. Every workflow has gates; the gates and the evidence vary per workflow.

The Learning Validation Pipeline is a sub-capability of Cross-Workflow Coordination in the Operating Layer. It takes outcomes and feedback from completed workflow runs and validates what gets promoted into the Intelligence Layer as new learned intelligence. Learnings are classified as validated, provisional, or discarded. The Worker decides what enters the body of knowledge; the Operating Layer prepares the evidence.

Every time a workflow runs, the Intelligence Layer captures what worked and what did not. Validated learnings promote into the Intelligence Layer. The next run of any workflow draws from richer intelligence. Over time, task execution improves because agents have more knowledge to draw on. This compounding is automatic and structural; it is not a feature that requires manual effort.

Cadence workflows loop: CRM nurture, performance reporting, and media optimization keep running on a schedule. Learned intelligence feeds back between runs, and the next run is smarter. Project workflows terminate: Brand Architecture Build and GMMM Build finish when the deliverable is produced. Learned intelligence still feeds back for adjacent workflows. Whether a workflow loops or terminates is a property of the workflow, not a per-run Worker decision.

Agents are AI workers that perform tasks inside workflows. They carry domain expertise and are governed by the Operating Layer. The Worker is the human at the boundary. The Worker authorizes, reviews, redirects, and judges, but does not perform tasks, govern agents, evaluate agent output, or interrupt workflows mid-task. Agents do the work; the Worker bounds them.

Capabilities organized marketing work by competence domain (brand, measurement, creative). Archetypes organize by engagement pattern (what the buyer asked for). Mark's direction: clients buy end-to-end experiences (acquisition campaigns, media optimization programs), not competence domains. Every archetype crosses multiple competence domains; the competence lives in agents and functional intelligence, not in a named layer.

The eight archetypes were derived from Mark's direction to organize AOS around what buyers ask for. They cover the major marketing engagement patterns a CMO buys from an agency: acquiring customers, retaining them, building a brand, optimizing media, producing content, executing campaigns, activating events, and measuring impact. The count is locked at eight because it covers the full marketing scope without overlap or gaps.

The Context Layer was added in 1.4 to make explicit how the same archetype adapts for different engagements. B2C acquisition and B2B acquisition are the same archetype; what changes is the context (B2B/B2C, Industry, Geography, Company Size). The Operating Layer's Intelligence Assembly draws different intelligence for different contexts. Without the Context Layer, the architecture implied that each context variation was a separate archetype.

The workflow shape framework (Build vs Run macro-shapes plus six operational patterns underneath) was retired in Session 8 because it did not earn its complexity at the strategic altitude. Workflows are defined by their archetype, their tasks, and their outputs. Adding a shape taxonomy on top created an extra classification layer that made the architecture harder to explain without adding clarity for the buyer or the operator.

"Substrate" is passive (something sits on it). "Engine" is active (something runs because of it). The Operating Layer and the Intelligence Layer drive workflow execution; they do not sit passively underneath it. The rename matches the "powers everything" framing the architecture uses. Conceptually unchanged; vocabulary modernized.

Session 7 collapsed the Worker's roles to three modes (Provide context, Authorize, Judge). Session 9 re-expanded because three modes compressed too much: Review (binary go/no-go), Redirect (change direction), and Judge (taste call) are operationally distinct moves the Worker makes at gates. Naming them separately is more honest about what running AOS actually feels like. The five-role frame maps cleanly to the original baseline without re-opening a separate Compound phase.

In 1.3, "AI and Agentic Operations" was listed as a capability alongside Brand, Measurement, and others. It was cut because AI is not a marketing domain at the same altitude as those competence areas. The work it described (agent governance, agent evaluation, brand code enforcement) was absorbed into the Operating Layer as machinery. In 1.4, agents are a named architectural concept, but the competence they carry is organized per task type in functional intelligence, not in a separate AI capability.