CBM v2 Cell — Agent Strategy

The AI-agent workforce is a first-class function in GrowDirect's operating model. It has a roster, a governance layer, a memory infrastructure, and a distinction between customer-facing agents and delivery agents.

Why this is a cell

Pre-2025 CBMs placed "IT tooling" under the Technology cell. Agents aren't tooling — they're labor. They do work, produce artifacts, collaborate with humans, and need governance. The moment agents become load-bearing, they earn their own cell.

Agent populations

Customer-facing agents

Agents that Canary Retail operates on behalf of a merchant, visible to the merchant's end users or operators.

• Chirp rule-evaluation agents (detection pipeline)
• Fox case-management agents (case triage, evidence collation)
• Owl search agents (natural-language retrieval over the merchant's canonical data)
• RaaS namespace-resolution agents (principal identification across the MCP mesh)

Delivery agents

Agents that run alongside the CATz delivery team during an engagement. Not visible to the merchant as "agents"; they appear as the delivery team's capability.

• Data Detective (../method/roles/data-detective) — source-to- canonical translation
• Digital Plumber (../method/roles/digital-plumber) — integration wiring
• Architect (../method/roles/architect) — target architecture production (planned)
• Writer (../method/roles/writer) — engagement documentation (planned)

Platform agents

Agents that keep the platform running, not visible to merchants or engagement teams directly.

• Memory-bus indexer
• Embeddings-rebuild scheduler
• Skill registry validator
• Atlas knowledge-graph maintainer

Memory-bus architecture

Every agent reads from and contributes to a shared memory bus backed by pgvector. The memory bus carries three kinds of content:

1. Curated knowledge cards — structured method fragments, data-shape patterns, integration patterns. Written by humans, read by agents.
2. Session memories — per-merchant, per-tenant, per-agent recall over prior sessions. Produced by agents, consumed by successor agents.
3. Decision records — architectural decisions, disposition calls, judgment-call resolutions. Authored by humans, recalled by agents when context matches.

Every memory chunk carries principal scoping. A delivery agent for one engagement cannot recall memories belonging to another engagement unless the partner explicitly cross-shared.

Principal-awareness — a hard requirement

Every agent in the GrowDirect mesh knows:

• Its own identity (which agent, which version).
• Its principal (whose authority it is acting under — the merchant tenant for customer-facing agents; the engagement team for delivery agents).
• Its authority scope (which MCP tools, which data, which actions it is authorized to invoke).
• Its audit trail (every action is hash-chained for evidence).

Agents that don't know their principal get exploited in an open- agent world. Principal-awareness is a security property, not a convenience.

Agent SDD template

Every production agent has an SDD covering:

• Purpose and scope
• Principal and authority boundaries
• Tools and data sources it reads / writes
• Memory bus interaction (reads what, writes what, retention policy)
• Failure modes and escalation paths
• Versioning and rollback
• Audit footprint

The template lives at method/artifacts/agent-sdd-template.md (to be populated).

Cross-cell relationships

• With Data Protection & Governance: agents carry PII in their context windows; PII handling flows through the DP&G cell.
• With PMO: the agent roster is part of the capacity model.
• With ARB: every agent addition or authority expansion is an ADR.

Related

• overview
• data-protection-and-governance
• arb
• ../method/roles/data-detective
• ../method/roles/digital-plumber