The three parts of a company brain.

The vision of the "AI-first company" is rapidly solidifying: an enterprise operated by autonomous AI brains, a living digital repository of company knowledge, workflows, and memory, maintained by agents first and humans second.

Y Combinator made this explicit in their Summer 2026 Request for Startups with the Company Brain category, and YC CEO Garry Tan open-sourced gbrain in April 2026, demonstrating an always-on agent daemon with 150,000+ synthesized pages running his personal and venture operations.

As founders rush to build or adopt company brains, a dangerous architectural mistake is emerging: treating all enterprise memory as one uniform problem, and therefore one unified solution.

Just as the human brain separates sensory perception, associative memory, and executive motor control into distinct regions, an autonomous company needs several distinct parts. Here I introduce split to three: a Rule-Driven Brain that decides, a Vibe-Driven Brain that senses, and an Operational Brain that holds the money, the records, and the machines.

Tier 1: the Rule-Driven Brain (operational conscience)

On the ground floor, running a business is not a vibe. It is a legal, contractual, and mathematical reality.

If your tax filing has a €0.50 discrepancy, the tax man does not care about your vector embeddings. If an invoice calculates utility meters using probabilistic LLM generation instead of arithmetic, you are committing billing fraud. If an autonomous agent signs an agreement without board authority, it is legally void, or it exposes you to personal liability.

This is the domain of the Rule-Driven Brain, the company's prefrontal cortex.

• The engine: Git-level version control, deterministic code, explicit state machines, and CI-style evaluation suites.
• The interaction: governed execution. The brain classifies work into a Control Plane (changing rules, granting permissions) and an Operations Plane (executing recurring business tasks).
• The output: exact calculations, audit trails, verified registry records, and drafted actions.
• The boundary: zero autonomous execution on external systems. The brain can read telemetry, calculate bills, and stage draft invoices or payments, but sending money, signing contracts, or submitting official filings requires an enforced human-in-the-loop (HitL) checkpoint.

---

Tier 2: the Vibe-Driven Brain (sensory radar)

At the 30,000-foot level, a company is bombarded with continuous, unstructured signals: calendar invites, meeting transcripts, Slack banter, investor emails, pitch decks, market research, and half-baked product ideas.

This is the domain of the Vibe-Driven Brain.

• The engine: fuzzy associative retrieval, RAG, knowledge graphs, and ambient daemons. Tools like Garry Tan's gbrain, Hermes, OpenClaw, Notion AI, or native Google and Claude workspace connectors are on spot.
• The interaction: you ask high-level, open-ended questions. "What did our seed investors mention about Q3 pricing?" or "Who on our team last spoke with Acme?"
• The output: well-synthesized, cited prose, trend alerts, and relationship maps.
• The boundary: the Vibe-Driven Brain is an observer and advisor, not an operator. It generates meeting prep and surfaces strategic leads. It must never hold the keys to company bank accounts, signing authority, hiring/firing employees or tax filings.

---

Tier 3: the Operational Brain (physical actions and accounts)

Physical sensing, accounts, and actions are not done with AI at all. Here you want something well-defined, reliable, and dependable: boring, old-school systems. Depending on scale and needs you can prefer systems internal to the company, or SaaS services. Just some samples:

• Operational production systems: ERP and related
• Sales and marketing, customer records and communication: CRM and similar
• Human resources, hiring and onboarding, holding personnel files and private data with particular care
• Accounting and inventory management
• Reports to shareholders and government
• Banking, finances, and assets
• Legal obligations for compliance
• Intellectual property: source code, databases, designs, documentation, contracts

---

The big mistake: replacing permission walls with system prompts

Many AI builders try to turn a vibe-driven system into an operator simply by writing a longer system prompt:

"You are a careful accountant. Never hallucinate, and never spend more than €1,000 without asking."

This is security theater. Prompts drift, models misinterpret context, and probabilistic engines hallucinate.

A true Rule-Driven Brain enforces safety at the infrastructure layer:

1. Hard API boundaries. Connectors to banks (like LHV) and cloud storage are hardcoded read-only. Accounting APIs (like Merit Aktiva) are hardcoded to allow draft creation only. The agent literally lacks the technical API scope to transfer funds or send invoices.
2. Deterministic math in real code. Utility meter allocations and VAT calculations are handled by unit-tested Python scripts, not LLMs. The LLM is invoked only to classify exceptions when calculations fail validation.
3. Auditability over vector databases. Key decisions and corporate policies live as Git-versioned Markdown files with machine-validated YAML frontmatter. Every change has a cryptographic author, a date, and a rationale, not an opaque vector weight.

---

The architecture in practice

For my own family microbusiness (a small asset rental company) the stack looks like this.

Tier 1: the rule-driven brain

• Operational brain lives in a private version-controlled Git repository, as a set of well-defined .md files, with issues as the TODO registry. It holds the corporate constitution, legal signing authorities, tenant lease registers, and workflow contracts. I use GitHub pull requests to review and to keep the history of all changes and decisions.
• The Merit Aktiva accounting system is connected via an API and MCP that I wrote myself, since none was available at the time (merit_api on GitHub). I run it in draft-only mode.
• For payment checks and reconciliation I use the very nice LHV AI MCP integration, which gives proper read-only access to current accounts.
• A lot of unstructured input, invoices and customer requests and the like, arrives by email, so I set up semi-automatic processing using a combination of Apps Script, parsing, and an agentic integration. The OCR-based CostPocket service parses PDF invoices from email, or photos taken in its app, and enters them straight into Merit. This has strong white-list based filters, so don't try to send me fake invoices - these will not get paid automatically.

The Git repository's folder structure is evolving, and currently lists as follows:

Area	Purpose
control/	Management and company-brain governance: instructions, authority, connectors, permissions, and controlled change
company/	Identity, governance, purpose, strategy, and durable policies
history/	Chronological record of verified events
decisions/	Decision records: context, owner, rationale, and consequences; one per .md file
operations/	Routine workflows, run controls, cadence, and recurring obligations
tools/	External-system catalog, permissions, and security rules
agents/	Agent roles, routing, scopes, and handoff protocol
projects/	One directory per finite initiative; _template/ starts a new one
registers/	Cross-company risk, obligation, asset, and contact registers
sources/	Evidence snapshots and source-quality notes
templates/	Reusable notes and records
inbox/	Unprocessed material; not authoritative until reviewed
archive/	Superseded or completed material retained for history

Tier 2: the vibe-driven brain

• The basic overlay runs on AI agents that can see the rule-driven brain alongside many other integrated datasets. I just use the standard ChatGPT, Claude, and occasionally Gemini, which see both of the other brains plus a lot of additional sources: Notion, raw email, Google Docs, the web, and whatever is already trained into the models.
• For longer-running analysis, keeping a laptop lid open is not that practical, so I use an agent runner in the cloud: a small, cheap virtual machine. It has always-on terminal control (tmux, herdr) and, inside that, several agent terminal UIs (TUIs): pi, which gives cross-provider model access, especially when connected to OpenRouter, plus the command-line versions of OpenAI Codex, Claude Code, and Gemini.
• The rule-driven brain ingests the most meaningful and actionable output of that chatter, by feeding structured memos into the Git repository's inbox/.

Tier 3: the operational brain, "old school" tooling

• High-density, high-concurrency transactional data (water and electricity meter telemetry, smart home actions, cost calculations for invoices) lives in a well-guarded custom PostgreSQL database system hosted on Supabase. This is the microscopic ERP.
• Official accounting and invoicing is done with specialized SaaS services: Merit Aktiva paired with CostPocket.
• If I had an active sales pipeline, I would consider Pipedrive, HubSpot, Zoho or similar. For now I have not needed any.
• For very simple operational use, even Google Sheets works as a flexible, if less guarded, online database.
• Some automations work without touching my system at all. My bank (LHV) has a "salary payment" feature which not only does all the needed tax and deduction calculations, but also posts the company's monthly income report directly to the Estonian Tax and Customs Board (EMTA). All I need to do is approve the salary payment in the online bank. The accounting system could post that report too, in addition to VAT reporting. Luckily, for a company this simple, Estonia does not require much more regular reporting than that.

The manual tier

The AI-first company still needs physical legs to do the walking, and that is left to me and my good family helpers.

• Daily and weekly operations: none.
• Monthly operations: the rental company sends out invoices, based on agreements and extra cost surcharge distributions. I have some old-school consumption meters that are not automated yet, and one heat pump (quite a new one, incidentally) with no usable API, so I need to check and enter about four numbers per month into the operational database.
• Occasional operations, maybe a few times per year: when rental tenants change, or providers change, there is recording of new contacts and contract details.
• The more involved yearly company report is still done by an accountant in the traditional way. I have tried to vibe-generate it, but so far it involves so many tricky decisions, even for my simple company, such as the assessment and classification of certain depreciations, that it has remained manual. The base data, like invoices, is already there automatically, so for me as company principal it is just a reminder that the report is due.

Use vibe-driven AI to see and listen. Build a rule-driven, Git-governed brain to think and decide. And never let your eyes drive the car without the prefrontal cortex having its hands firmly on the wheel.