Cortex IQ sells AI agents and uses AI agents to run itself. That's not a marketing detail, it's a working rule: nothing we deliver to a client goes live before it runs in our own operation first. There's an old saying that the cobbler's children go barefoot. Here, the shoe is the same one we forge for the client.
This piece shows how that works on the inside, with the numbers checked against our own repository folders on October 30, 2026, and the list of what's still incomplete. Anyone who works with AI inside a company knows both parts matter.
Why AI consulting can't run on a spreadsheet
A company that recommends AI agents, data governance and software architecture, while running its own day-to-day on a spreadsheet and WhatsApp memory, has a credibility problem. Worse: it has a knowledge problem. You can describe how an agent fails on a slide, but only someone who runs dozens of them knows where the judge's score drops, what it costs in tokens, and when the quota runs out.
That's why Cortex IQ adopted what we call cyborg architecture internally too: people set the direction and answer for the result, AI executes, scales and flags. The difference is that here there's a repository that proves the split.
The four pieces and how they connect
The repository keeps four areas in the same place, on purpose: the agent has to read the vault before working, and leave a trace in it afterward.
- Skills. 126 files of trade rules, organized into 18 roots (the orchestrator and 17 roles, like CEO, CFO, CTO, CMO and CLO) and up to three levels deep. The skill is the single source of the rules. Agents and the website point to it and never copy it. If the rule changes, it changes in one place.
- Agents. 61 of them, across six groups: engineering, marketing, C-level, council, operations and data. Each agent is a skill with its own context: persona, input/output contract, quality rubric, and the source skill it reads before starting. One file per agent, validated by lint.
- agent-core. The runtime. Claude generates the artifact, a Portuguese-language linter checks it for LLM/cliché phrasing and chatbot residue, and a second model, Gemini, judges the text against a rubric. The writer doesn't grade its own work. If it fails, the text goes back to the generator with the judge's feedback, up to three tries. There are 27 rubrics, all inheriting four base criteria: completeness, absence of hallucination, technical alignment and clarity. The minimum score is 8.5, and any single criterion below 6 fails the piece even with a high average.
- Second Brain. An Obsidian vault with four separate vaults: Cortex IQ itself and each product division (Assetly), each isolated from the others. The agent reads the company's context before generating anything. After approval, the orchestrator writes the log, draft and index, always appending. Nothing is overwritten, and a revised decision becomes a new entry.
How a task moves through the system
Marketing and engineering follow different chains, and no agent calls another one directly: a person or the orchestrator sets the sequence.
In the marketing chain, the channel analyst (website, LinkedIn, Instagram, email, press) produces the piece. The marketing manager triggers the brand-guardian as a quality gate. The CMO approves, approves with changes, or rejects, and never rewrites. A person is the one who publishes.
In the engineering chain, the user story ships with front-end, back-end and test tasks, and passes through a gate before technical analysis. The technical package (architecture, data model, flows, API contract and security) passes through another gate before any code is written. Only then are back-end and front-end built in parallel against the same contract, followed by PR review and a security review that runs both before and after the code.
Who decides
The human decides. A C-level agent's recommendation is logged in the vault as "awaiting human decision," never as a decision made. Reviewers are read-only and never fix what they review. The approver doesn't rewrite. Scheduled agents watch, but never take an action that changes the environment.
For strategic decisions there's also a council room: six lenses (logic, strategy, first principles, ethics, systems thinking and philosophy) read the same question in parallel, and a core synthesizes them. That result also waits for a human decision and only reaches the vault if a person marks it so.
There are six rules that apply to everything, and the most-used one is also the one people find most inconvenient: never invent. An unconfirmed fact gets flagged "pending validation." A number, a client or a date only goes in with a source.
The virtual office
To register, organize and run the agents, there's the Cortex Office: a React interface with an isometric 3D office, rooms by department, agents at their desks, and the day's governance indicators on the wall. You can hire a new agent through an assistant (linted before saving), chain agents into workflows that pause when the judge rejects a step, open the council room, and talk to each agent directly. In "investigate" mode the agent only reads. In "resolve" mode, every edit, command, or web access asks for approval first.
The Office runs only on the local machine, is never exposed on a network, and never deletes anything: a removed agent goes to an archived folder, and deletion requires typing a confirmation.
What still doesn't work
An account of an operation that only lists what works doesn't help anyone. In the review we did on October 30, these points were still open:
- The n8n automations folder exists, but it's empty. The first real automation hasn't been defined yet.
- The repository's own supporting docs disagree with each other on the skill and agent counts. The fix is to generate the count by script instead of keeping it by hand.
- Agents' image prompts still use the brand's old visual identity. The migration is planned.
- The bridge between episodic memory and the curated vault is incomplete, and the isolation check between divisions needs to come before turning it on.
- An in-progress workflow is interrupted if the local service restarts, and it runs one execution at a time.
It works, and the open items are known and written down in the vault. That's the same standard we hold clients to.
What this means for anyone hiring us
The same design becomes the deliverable. An AI agent in production needs a written trade rule, a clear input/output contract, independent evaluation, traceable memory, and a human with the final decision. Without that, the agent works in the demo and breaks on a Tuesday. That same skeleton of trace and decision answers, in practice, what AI governance asks for: who did it, based on what, who approved it.
If a company wants to start without reorganizing everything, the path we recommend is to pick one trade, not the whole company. Write that trade's skill, build an agent with a rubric, connect it to a vault, and measure for 90 days, the way we show in how to measure an AI agent's ROI. That first slice fits inside a PoC through Cyborg Forge, and the technical leadership to decide what comes next can come from CTO As a Service.
Frequently asked questions
What's the difference between a skill and an agent? A skill is the text with a trade's rules, like the rules for a CFO or a brand reviewer. An agent is a skill with its own context: persona, input/output contract, quality rubric, and access to the memory vault. The rule lives in the skill, and the agent reads it before starting.
Who approves what the agents produce? A person. In marketing, the piece goes through the brand reviewer and the CMO, and publishing is human. In strategy, the agent's recommendation is logged as awaiting a human decision, never as a decision already made.
Why use one model to generate and another to judge? Whoever writes something doesn't grade it well. An independent judge, with a rubric and the vault's context, catches what the generator misses, like an invented fact or an unsourced claim. In our setup, Claude generates and Gemini judges, with a minimum score of 8.5 and up to three tries.
Do the agents access client data? Only what's in that client's vault, in its own project folder. The vault never stores passwords, tokens, API keys, or signed contracts: it stores a summary and the Drive path. Product divisions are isolated from each other.
Can I build something like this at my own company? Yes, and the best place to start is a single trade, like support, sales, or marketing. Write the skill, build the agent with a rubric, connect it to a vault, and measure for 90 days. Running agents for clients requires the client's own account or API key.