At Zarlu, I’m building an agentic C-suite.
Right now, I have agents acting as CFO, CMO, CRO, CDO, COO, and General Counsel. Each has a narrow scope, limited permissions, and scheduled work. The bet behind this setup is simple: if LLMs keep getting better over the next few years, the companies that already know how to turn them into leverage will move faster than the ones that wait.
I’ve followed AI since the deep learning wave around 2011, when the best systems could do things like classify digits and tell cats from dogs. Since then, the models have become much more capable, and progress has not stalled for any meaningful stretch. I do not know exactly what models will look like in 2029. I do think they will be far better than they are today. So I want to build for where the puck is going, not where it is now.
That does not mean I want to disappear into internal tooling and forget the actual business. The meta work is not the business. In the last few months alone, the way I build automations has already changed at least twice. So I am trying to strike a balance: invest enough in internal systems to build leverage, but not so much that it distracts from getting products to market and finding product-market fit.
Here is what the current setup looks like.
The CFO has read-only access to financial data and reports on cash flow, expenses, and changes in spending. I plan to integrate QuickBooks next so that workflow can go deeper into accounting.
The CMO proposes SEO strategies and writes first drafts for blog posts across product sites.
The CDO connects to PostHog, checks how products are performing, and looks for instrumentation and observability gaps.
The CRO helps prepare sales leads using Apollo and Instantly.
The COO acts as the orchestrator, keeping the other agents on track and making sure work moves.
Some of these roles are further along than others, but even in their current form they are already useful. I use AWS Lambda functions to trigger agents on a schedule when I expect work to be done.
This is not autonomy in the sci-fi sense. These are supervised systems with bounded roles, constrained access, and clear blast-radius limits. They save me time each week by drafting content, monitoring telemetry, preparing leads, and surfacing things I would otherwise have to check by hand.
I also spend a lot of time on safety. I sandbox the agents and filter the external content they interact with to reduce the attack surface if a malicious prompt is injected or an agent goes off script.
I do not know if this exact structure will be the right one forever. The company will change, and the tools will change with it. But right now this model feels promising. Each week it gets a little more useful, which lets me spend more time talking to customers and getting products into the market.
I’m curious how many other founders are doing something similar. If you are, I’d love to compare notes.
John