Skip to content

Case study · hero · 2025 to present

Simple Cortex

An AI consulting practice run as a three-agent operating company, with cost as a first-class tool surface.

Role
Founder and operator
Company
Simple Cortex
Dates
2025 to present
Bucket
current venture
Simple Cortex homepage hero on a dusty mauve background: large condensed-display headline reading BETTER SYSTEMS BETTER BUSINESS., body copy We help small businesses improve operations, reduce inefficiencies, modernize their systems, and adopt the right technology without unnecessary complexity, and a logo strip We've helped amazing companies like Microsoft, Samsung, Salesforce, Verizon
Simple Cortex · HomepageThe consulting practice positioning. One sentence, one customer-logo strip, one promise. Better systems is the offer. Better business is the receipt.

TL;DR

One-line outcome. I built and operate a three-agent AI consulting practice on a local-first model router, so the company ships client work without burning consulting margin on cloud inference.

Role, dates, scope. Founder and operator, Simple Cortex (Northern Virginia, simplecortex.com), 2025 to present. I own company strategy, agent persona definitions, the OpenClaw Router rules, and the public client surface. Active client engagements with Kintsu Medspa, Safe Driving Academy, and eCom Logistics.

Metric chips.

  • Three named agents in production, each defined by a four-file contract (AGENTS.md, SOUL.md, HEARTBEAT.md, TOOLS.md) committed to git.
  • Six semantic routes on the OpenClaw Router, local-first with cloud fallback reserved for judgment calls.
  • Cost visible by default. ocr costs is a daily-driver command, not a billing-dashboard item.

Key decisions

Agents

Router

Models

LOCALCLOUD FALLBACKCEOCALLS: ARCHITECTFounding EngineerCALLS: ENGINEERResearcherCALLS: ANALYSTOpenClaw Routerqwen3.5:9bCOMPANIONqwen3-coder-next:q4_k_mENGINEERopenclaw-qwen35-a3b-thinkANALYSTqwen3.5:4bTRIAGEllama3.2:3bCLASSIFIERclaude-opus-4-6ARCHITECTclaude-sonnet-4-6SENIOR, JUDGECLASSIFYROUTEFALL BACKMETER COST
Agents call by intent. The router classifies, routes, and falls back. Cost is metered on every call.

Four files per agent, not one. The cheap path was a single agent.md per agent with persona, instructions, tools, and execution checklist all together. I split it into AGENTS.md (instructions), SOUL.md (persona and voice), HEARTBEAT.md (execution checklist), TOOLS.md (tool surface). Persona is now reviewable independent of instructions. The tool surface is reviewable independent of the execution loop. A new operator can read any single file without paging through the others. Information architecture applied to agent definition.

Router-first, model-blind agents. Agents do not name models. They call the OpenClaw Router by semantic intent: code, research, organization, conversation, visual, architect. The router classifies first, routes second, falls back third. The bet: cost discipline and model portability are worth more than running clean against a single provider.

Cost as a tool surface. ocr costs is a daily-driver command, not an admin curiosity. Provider health and routing decisions are first-class endpoints on the router. The operator sees spend the same way they see uptime, not at quarter-end.

Voice codified as a leadership artifact. I wrote the CEO SOUL.md as a leadership philosophy document, not a chatbot prompt. "Default to action." "Treat every dollar, headcount, and engineering hour as a bet." "Be replaceable in operations and irreplaceable in judgment." Voice lives in git and gets reviewed like code.

What changed

Agent / CEO

Four files, one contract

AGENTS.mdInstructions

What the agent does, in plain prose.

Run the company. Brief the operator. Decide.

SOUL.mdPersona

Voice, values, decision style.

Default to action. Treat every dollar as a bet.

HEARTBEAT.mdExecution loop

How a session runs end to end.

Open the brief. Decide. Log the decision.

TOOLS.mdTool surface

Which MCP tools and CLIs the agent can call.

ocr route. ocr costs. git. read repo.

Four files per agent. Persona is reviewable independent of instructions. The tool surface is reviewable independent of the execution loop. Information architecture applied to agent definition.
  • Three named agents (CEO, Founding Engineer, Researcher) are committed to git with a four-file contract each. Any new operator can audit any agent in four files instead of one monolithic prompt.
  • The OpenClaw Router ships with a CLI (ocr), a full README, and an HTTP API. Six semantic routes, local-first, with cloud fallback reserved for the cases that need it.
  • Agent memory directories are gitignored, so persona and tacit context persist locally but never enter the public repo.
  • The Founding Engineer agent loads a named skill stack explicitly (/investigate, /qa, /review, /ship, /design-review, /freeze, /guard). The workflow is codified: investigate before fixing, review before shipping, QA before marking done.

Measurable outcomes

  • Public landing page live at simplecortex.com.
  • Three named agents in production, each defined by a four-file contract committed to git.
  • OpenClaw Router shipped with CLI, README, HTTP API, and six semantic routes, local-first with Claude Sonnet 4.6 and Opus 4.6 as fallback.
  • Active client engagements: Kintsu Medspa, Safe Driving Academy, eCom Logistics.
  • Cost-per-task surfaces on the router's cost dashboard, so the operator can see spend by route in real time, not at quarter-end.

Leadership lens

I run Simple Cortex the way I would run a design org. Write the operating system before delegating. Treat voice as a system that can be reviewed and revised. Make business telemetry visible by default. The persona files are not prompts. They are leadership artifacts. The router is not plumbing. It is where the company's cost discipline is encoded.

What I did with my hands

Player-coach proof. I author the operating layer, not just the strategy.

  • CEO `SOUL.md` end to end. Principles ("Default to action," "Treat every dollar, headcount, and engineering hour as a bet," "Be replaceable in operations and irreplaceable in judgment"), hard rules, and the exclamation-point ban. Voice is versioned in git and reviewable as a leadership artifact, not buried in a prompt.
  • OpenClaw Router CLI (`ocr`). I wrote the primary operator surface for routing previews, cost reports, provider health, and stats. ocr route "<query>" lets a developer preview routing before committing. ocr costs is a daily-driver command.
  • Four-file agent contracts. I wrote the AGENTS.md / SOUL.md / HEARTBEAT.md / TOOLS.md split for each agent. Persona reviewable independent of instructions. Tool surface independent of execution loop.
  • Founding Engineer skill stack. I codified the workflow as an explicit named-skill list (/investigate, /qa, /review, /ship, /design-review, /freeze, /guard) so the loop is fixed: investigate before fixing, review before shipping, QA before marking done.

Reflection

A consulting practice that bills clients while paying retail for AI is a thin business. A practice that runs lean on local models can sell those hours back to its clients instead. At Director scale I would do two things differently: pull the persona files into a shared style guide so every new agent inherits voice rules (the same way a design system enforces token discipline across components), and add a lightweight UI over ocr so non-engineer operators can audit cost and routing without the CLI.

_Stack colophon: local Ollama tier (classifier, companion, engineer, analyst, vision) with Anthropic cloud fallback for senior and architect roles. Paperclip with the claude_local adapter as the agent runtime._