--- title: The Cognition Plane url: https://revworx.ai/architecture/8/ date: 2026-10-20T14:59:02+00:00 modified: 2026-11-08T03:24:46+00:00 lang: en_US --- Section 8 · Part two # The Cognition Plane The Cognition Plane decides what happens. Its first design commitment concerns the people who use it: your operators speak to the system, not to individual tools. This is the interface revolution of Section 2 made concrete; Section 12 describes the operator surface in full, along with the other surfaces through which the same engine meets staff and the public. A weekly pipeline review, approval of outreach copy and a request for a new integration all happen in one conversation, and the system works out which capability each request needs. In this section [8.1What belongs in the supervisory layer](#8-1)[8.2The Context Engine](#8-2)[8.3Specialists and capabilities](#8-3) Part two [§7The shape of the architecture](/architecture/7) §8The Cognition Plane [§9The Tool Plane](/architecture/9)[§10The Execution Plane](/architecture/10)[§11The Data Plane](/architecture/11) [Download the PDF →](/wp-content/uploads/2026/10/composable-reference-architecture.pdf) In brief The Cognition Plane decides what happens. Your operators speak to the system, not to individual tools, and it works out which capability each request needs. Everything that must hold across all work sits in one supervisory layer, and capabilities are roles inside one engine. Generic output is an architecture gap, not a prompting problem. Underneath, the plane divides into a supervisory layer and a set of specialists. The supervisory layer, which we call Agent Main, holds everything that must be true across all work: routing, context, guardrails, approvals, measurement and memory. The specialists are sub-agents, each with its own instructions and its own limited set of tools. People meet the whole through agent surfaces, the conversational interfaces described in Section 12. ![Figure 5 — The Cognition Plane: supervisory layer and specialists, with capability counts](https://revworx.ai/wp-content/uploads/cra-figure-5-cognition-plane.png) *Figure 5 — The Cognition Plane: supervisory layer and specialists, with capability counts* Current implementation The supervisory layer's operator-facing surface is the GTM Agent, through which operators direct and oversee the five sub-agent groups shown in Figure 5. The sub-agents are internal; the GTM Agent is where their work is directed and seen. ## 8.1What belongs in the supervisory layer Seven components, each there because the alternative is repeating it in every specialist. The **Supervisor** routes work by capability, task and context. It is a component of the supervisory layer, not an agent people address; the agents people address are the surfaces of Section 12. The **Context Engine** holds the operating knowledge of your business. **Guardrails** enforce brand, compliance and operational limits, including a suppression-list check before every outbound action; a rule like that must live in one place or it will eventually live in none. **Human-in-the-Loop** routes the decisions that stay with people: outreach copy, proposal sign-off, suppression disputes, partner payouts. The **KPI Monitor** watches your metric set continuously and raises threshold alerts rather than waiting for a reporting cycle. The seventh component, **Just-in-Time Context and Vector Memory**, carries both names for the reason given in Section 6. By default the agent fetches the record, transcript or analytics slice it has decided it needs. The vector store remains for recalling meaning across past conversations and documents. ## 8.2The Context Engine The failure revenue leaders most expect from agentic GTM is generic output: copy that could be about any company, in any category, to any buyer. The Context Engine prevents it, and it is the component most often missing from otherwise competent agent architectures. It maintains four libraries and supplies them to every sub-agent while it works. Three establish who your buyer is, what story your brand tells them and which evidence they find credible. The fourth is different in kind: a graph linking each revenue outcome to the drivers that move it and the gates that block it. That graph is what turns measurement into action. Without it, a KPI alert is a notification rather than a next step. ![Figure 6 — The Context Engine: four libraries and the learning loop that refreshes them](https://revworx.ai/wp-content/uploads/cra-figure-6-context-engine.png) *Figure 6 — The Context Engine: four libraries and the learning loop that refreshes them* For technical evaluators All four libraries are plain markdown under version control, so a change to how the business describes itself is a reviewable diff rather than an untracked prompt edit. The learning loop proposes updates from transcripts and KPI observations. Those updates are commits, so they can be reviewed, reverted and traced to the evidence behind them. The Revenue Lever graph refines its edge weights as outcomes accumulate. ## 8.3Specialists and capabilities Five sub-agents carry twenty-eight capabilities between them. A capability is a role inside the agent engine, with structured instructions, explicit output expectations and a limited set of tools. It is not a separate process, and it is not a separate subscription. That is the economic argument in one sentence. Because capabilities share infrastructure, memory, context and guardrails, adding the twenty-ninth costs a set of instructions and permission to use a tool, not a vendor evaluation, a contract and an integration. It is also why one operator can run all of them. What composes hereThe model behind the reasoning, individual capabilities, and the contents of the four libraries. Each can change without touching the tools, the infrastructure or your customer record. Current implementation - **Sales Execution (5)** — proposal and quote drafting, deal risk analysis, opportunity hygiene, competitive intelligence, conversation intelligence. - **Demand & Intelligence (6)** — intent signal monitoring, research and segmentation, waterfall enrichment, partner-portal crawling, lead routing, lead and account scoring. - **Engagement & Outreach (9)** — outreach orchestration, reply triage, content and copy, meeting booking, e-commerce, storefront operations, venue concierge, paid media management, dynamic nurture. - **Success & Expansion (4)** — churn risk detection, customer health monitoring, quarterly business review preparation, upsell and cross-sell. - **Cross-Cutting Services (4)** — knowledge base, compliance and privacy, attribution analysis, forecast and pipeline analysis. Key takeaway Put everything that must hold across all work in one supervisory layer, and treat capabilities as roles inside one engine. Generic output is an architecture gap, not a prompting problem. Cite as: revWorx (2026). A Composable Reference Architecture for Autonomous GTM Operations, First Edition, §8. revworx.ai/architecture/8