Agentic Brand Operations: When the Back Office Runs Itself

Some of the most interesting thinking gets smoothed out in the editing process. The hedged answer becomes the clean takeaway. The half-formed idea that was actually going somewhere gets cut for length. We've noticed this for a while, and this interview is our attempt to do something about it. What follows is a conversation with Pavan Otthi, founder of Curator — on why commerce operations is one of the hardest automation problems in software, what's actually new about this generation of agents, and what it takes to build something founders can actually trust to run.

Here's what we heard.

Current Business Context

UpScaleX: What was the original pain point that led you to build Curator?

Pavan: Disconnection. We've spent the last couple of years building in agentic commerce on both the B2B and B2C sides. We started by focusing on how fragmented the product discovery journey was for shoppers, but working closely with brands revealed a deeper problem: the operators themselves were stitching together everything from supply chain to marketing across a sprawl of disconnected SaaS tools that didn't talk to each other. Curator is how we envision work being executed in the future, where lean operators stay lean and agents handle the scale.

UpScaleX: What does "agentic commerce operations" actually mean in practice, and what does it not mean?

Pavan: It doesn't mean AI replacing the people who build the brand. I'm bullish on AI taking over the manual, repetitive work of running a business, but taste and creativity aren't going anywhere. Agentic commerce operations is what happens when operators get to spend their time on the next chapter of the brand instead of digging through dashboards to figure out why a shipment got delayed.

UpScaleX: Why is commerce operations a uniquely difficult automation problem?

Pavan: Sensitivity and variability. Each brand uses its own cocktail of SaaS tools and has its own spin on how to best operate as a business. On top of that, day-to-day operations are crucial, and one mistake in automation could be very costly. Having the proper approval modes, auditability, and observability is crucial in building an automation platform for commerce operations.

UpScaleX: How do you define operations automation?

Pavan: Operations automation is giving agents control of data entry, constant monitoring of the business, repetitive tasks, and insights. It's putting the operator in the driver's seat and not responsible for refilling the tank and oiling the engine.

UpScaleX: What's actually new about agentic systems today, versus just smarter automation or better dashboards?

Pavan: Reasoning, omnipresence, and action ability. Today, agents can use your browser, keyboard, and mouse to perform the same actions humans do on their computers. That lets us operate across platforms that don't offer clean API integrations. In addition, ambient AI, or, in other words, AI that's constantly watching how operators make decisions and every moving piece of the business, allows for dashboards to truly be living, breathing windows into every aspect of the business.

The Personal Assistant Frontier

UpScaleX: When was the ah-ha moment for you?

Pavan: The realization that agents could actually do some of my daily work. It felt like we were getting 2x or 3x the productivity we could have in a day, which is pretty crazy.

UpScaleX: Tell us about your side project, Jarvis. How did it come about, and what surprised you about the current architecture?

Pavan: This was a fun one. Recently we've been super busy, so we haven't had much time to work on it, but we're still using it to run some X and LinkedIn engagement.

We started by forking OpenClaw and building out infrastructure for better computer use, so we could have it run for longer. Eventually we used its core open-source agent runtime called Pi, which is honestly the unsung hero behind OpenClaw. It's a really dynamic and customizable agent framework that comes with coding tools and branching out of the box, which allows it to adapt and overcome challenges that would stop long-running agent tasks otherwise.

We're using it for computer-use tasks like navigating our X feed, finding interesting posts on LinkedIn, and identifying potential customers. Right now we're focusing on releasing it as an outbound, go-to-market, personal branding agent, as well as a personal assistant.

Beyond scripted agents or back-and-forth chat, there's truly been a shift in the agent world toward focusing on agent harnesses, how we enable agents to perform dynamic, long-running tasks with complex reasoning to truly unlock value for their users.

How This Evolves

UpScaleX: What does this mean for business and operations teams going forward?

Pavan: LLMs will keep getting better, which means users will be able to trust their agents to take on more audacious tasks. Teams should focus on staying lean, and individuals should become generalists — so they can delegate specialized work to specialized agent environments.

UpScaleX: What will "good" look like, and how will we evaluate the complex outputs of agents?

Pavan: Agent harness eval is definitely a hot topic right now, and I'd say it's still pretty nascent. Some say the top labs are overfitting their models to perform well on tool-calling and agent-eval benchmarks. One of the things we're really focused on is building a bunch of golden query sets to perfectly mimic how our customers would use their agents in the wild, and focusing on accuracy, latency, and reliability in dynamic data environments.