Inside the Agency Knowledge Base
Most AI agents fail because there’s nothing structured underneath them to read. This is the layer Brain in a Vat Music Promotions built: a knowledge base agents read before doing anything.
The Shape of It
The knowledge base lives alongside the agency’s working files. Four sections of single-purpose pages — one concept per page, cross-referenced throughout. Business explains what the agency is and how it’s run. Functions explains what it does. Capabilities explains how it does it. Foundation explains the conventions everything else keys off.
- Brand voice & visual identity
- Strategy
- Legal entity, banking, insurance
- What it is, who runs it, where it operates
- Campaign execution
- Deal flow
- Financial
- Operations
- Service development
- Asset management
- Pitch list generation
- Pitch send infrastructure
- Payment & invoicing
- Automation rules
- Scheduled workflows
- Schema conventions
- Naming conventions
- Billing entity conventions
- SSoT principles
- Taxonomy
- Decision log
What a Page Looks Like
Every page has a structured header (what it’s the source of truth for, what it relates to), section headings, body content, and cross-references at the bottom. The structure is consistent, so an agent reading one page knows where to find things on every other page.
Naming Conventions
knowledge-base / 4 Foundation / Naming conventions.md
Airtable Field Names
- No parentheses in new field names —
Total X, notX (Total). - Singular over plural for entity fields —
Client, notClients. - Cross-referenced fields prefix with the source table when ambiguous —
Campaign Clientinstead ofClienton a related table.
Script and File Names
- Kebab-case for shell and node scripts —
enrich-campaign-metadata.js. - Snake_case reserved for Python only.
How Agents Use It
The knowledge base has an index that maps each type of work to where it starts. Email feedback interpretation, asset chasing, SubmitHub ingest — there’s a playbook for each one. The agent reads it first, then acts. When something goes wrong or a new edge case comes up, it gets logged. The next agent working that same job doesn’t have to figure it out again.
