The Problem the Brain Solves
Foundation models (the GPT-class large language models from OpenAI, Anthropic, Google, Meta, and a few others) are general-purpose. They know roughly everything that was on the public internet at training time. They do not know your business.
Your business has: a specific customer list, a specific product catalog, a specific way you talk to customers, a specific set of tools you use, specific rules about what you do and do not say, specific accounts payable and accounts receivable, specific deals in progress, specific support tickets open. None of that is in the foundation model.
To turn a foundation model into a useful operator for your business, you have to surround it with the layer that holds all of that context, gates the model's actions, learns from what worked, and stays current as your business changes. That layer is the brain.
Without the brain, an AI agent is a foundation model with vibes. It can write plausible-sounding emails but it does not know your customer's history, does not know which deal it is on, does not know your tone, and cannot take real action in your tools. With the brain, the same foundation model becomes an operator that knows your specific situation and acts inside the rules you set.
What Is Actually in a Business Brain
Every serious vendor's substrate has roughly the same eight components. The names differ; the components do not. This is convergent design - the problem fixes the architecture.
**1. Knowledge graph.** A structured representation of your business as entities and relationships. Customers, products, deals, vendors, contracts, playbooks. Not a pile of text that the agent searches; an actual graph with typed nodes. Garry Tan's GBrain calls this out explicitly; Sentie ships a per-tenant Business Brain.
**2. Skills registry.** A catalog of executable actions the agent can take. One file per skill, each with the SDK reference for the underlying tool, the input contract, and the success condition. Send an email. Create a deal. File a ticket. Move money. The skills registry is what gives the agent its hands.
**3. Memory layer.** Per-tenant persistent context. What the agent has tried, what worked, what your team specifically wants kept or kept private. The thing that lets the agent get smarter about your business over months without you re-teaching it.
**4. Hybrid retrieval.** Vector search plus keyword search plus graph traversal. Modern brains do not bet on one retrieval strategy; they ensemble. When the agent needs context, it pulls via the right method for the question.
**5. Model routing.** Which foundation model handles which task. A cheap fast model for classification (categorize this email). A frontier model for hard reasoning (should we offer this customer a refund?). A voice model for speech. The brain owns these decisions per-request so you are not paying frontier-model prices on every keystroke.
**6. Durable job queue.** Long-running agent tasks need to survive process restarts, retry on failure, and persist their state. Most brains use Postgres or similar. GBrain calls this layer Minions. The pattern matters because real customer-facing agents take hours or days to complete some workflows.
**7. Voice interface.** Voice in, voice out, tied to the same brain the text interface uses. Increasingly default rather than optional; most real-world buyers want phone-call coverage from the same agent that handles email.
**8. Constitution file.** A plain-text markdown spec that defines defaults, tone, throttles, what the agent must never do. The agent reads this on every run. Karpathy's CLAUDE.md is the most famous example; Anthropic Skills use the same pattern; Sentie ships a tenant-specific version.
Not every implementation has all eight. The serious ones aim at all eight and own up to what is missing. Vendors that talk about their agent without ever describing the substrate are usually selling you the foundation model with marketing on top.
Why the Same Pattern Keeps Showing Up
Five separate teams arrived at this same substrate architecture in the last six months without coordinating:
- Anthropic shipped [Claude Managed Agents and Skills](https://www.anthropic.com/news/claude-managed-agents) - Andrej Karpathy released a CLAUDE.md repo that hit #1 on GitHub at 44K stars - Garry Tan open-sourced GBrain (his "opinionated agent brain") with knowledge graph + skills + Minions queue + MCP - Eric Siu / Single Grain open-sourced 17 categories of marketing agent skills - Sentie shipped the Business Brain as the middle layer of our three-layer stack
None of these teams talked to each other. The pattern is convergent because the problem fixes the architecture. Once you accept that you need to turn a foundation model into a custom operator for a specific business, the eight components above are what you end up with.
The practical implication for buyers: when a vendor cannot tell you what is in their substrate, they probably do not have one. They are selling you a foundation model with a frontend.
How Sentie's Business Brain Works (Dogfooding)
Sentie's stack is three layers. The Business Brain is the middle layer.
1. **Agent layer** (what the customer experiences). Core agent + specialized agents. The thing your team interacts with. 2. **Business Brain / substrate** (what makes agents custom). All eight components above, per-tenant, isolated. 3. **LLM Surface** (how external systems connect). [REST API at api.sentie.io/v1](/developers/api), [sentie CLI](/developers/cli), and a [public OpenAPI spec](/openapi.json).
The Business Brain is configured during the free assessment phase. Your dedicated Success Manager runs discovery, captures your products, customers, playbook, voice, and tools, and builds out the substrate before any agent ships. The substrate gets richer over time as the agent runs and learns; this is the compounding part of the value proposition.
We use the same substrate internally. Sentie's own sales follow-up, customer onboarding, content publishing, and operational support run on Sentie agents backed by Sentie's own Business Brain. We do not ship a substrate capability to customers until we have proven it works on ourselves. The dogfooding is the credibility layer.
What to Look For When Evaluating a Custom AI Agent Vendor
If you are buying a custom AI agent for your business, ask these six questions. The answers tell you whether the vendor has a real substrate or is selling you a foundation model UI with marketing.
**1. Is your data stored as a knowledge graph or just a vector store of documents?** Knowledge graph means the agent understands your business as entities (customer, deal, product) with relationships. Vector-store-only means the agent searches text. Both have a place, but a serious substrate has both.
**2. Does the agent route across multiple foundation models or is it locked to one vendor?** Single-vendor lock means you pay that vendor's prices on every task and inherit their outages. Multi-vendor routing means the right model for the task at the right price.
**3. Show me your skills registry.** How many tools does the agent already know how to use? How fast can a new integration be added? A real substrate has dozens to hundreds of skills ready out of the box. A foundation model with a frontend has none.
**4. How does the agent persist state across long-running workflows?** If the answer is "it does not, each task is independent," the agent cannot do multi-step work that takes hours or days. Look for a durable job queue (Postgres-backed or similar).
**5. Can I read the agent's constitution file?** The plain-text spec that says what your agent will and will not do. If the vendor cannot show you this file or it is hidden in a dashboard you have no edit access to, the agent is opaque. Run.
**6. Is the substrate accessible via an API?** Mature substrates expose themselves programmatically so your engineering team or external AI assistants can interact. Closed substrates create lock-in even when the vendor swears they do not.
At Sentie, all six answers are yes-with-detail. If you are evaluating us, ask. If you are evaluating someone else, ask them.
Where the Pattern Goes Next
Three forward-looking trends worth tracking if you are paying attention to this space:
**Agentic commerce.** Stripe shipped a [Link wallet specifically for AI agents in May 2026](https://link.com/agents). Agents can now hold a wallet, request one-time-use cards, and execute purchases with human approval per transaction. The substrate now needs a wallet component, not just an integration. Marketing-spend agents, procurement agents, parts-ordering agents become viable categories.
**Open standards for agent-to-agent connection.** Protocols for how AI assistants discover and invoke external services are crystallizing. Brains that publish a documented developer surface (REST API, OpenAPI spec, CLI) become discoverable to LLM clients on the buyer's behalf as those clients add support for the relevant protocols. Brains that hide behind closed UIs are not.
**The .md file as the agent programming language.** CLAUDE.md, AGENTS.md, DESIGN.md, SKILL.md. Plain-text spec files that AI assistants read at runtime. The format is winning over JSON configs because foundation models read prose better than they read structured data, and humans can review markdown without engineering help.
If you are evaluating a vendor today and they have not heard of any of these trends, they are not building for the next two years.
The One-Sentence Summary
A Business Brain is the substrate underneath a custom AI agent - knowledge graph plus skills plus memory plus retrieval plus model routing plus a job queue plus a voice interface plus a constitution file - and it is the part of a custom AI agent vendor's stack that actually makes the agent custom to your business.
If you want to see Sentie's substrate up close, the [free assessment](/onboarding) is the fastest path. Your Success Manager walks you through what your Business Brain would actually hold for your specific business. No credit card, no commitment.