The AI that said No.

Published on January 19, 2026 by Raff Paquin

Imagine the scene: I've been thinking about a new feature all weekend for my budgeting app.

Finally, Monday morning. I'm excited about the idea and ready to build. I work a good 20 minutes to fine-tune the specs, and the issue is ready: R-648 — Add support for Plaid investment accounts.

Next, in Linear, I assign the ticket to my Obi agent.

Obi joins the session, spins up a cloud instance to checkout the latest version of the code, and assigns the ticket to my product management squad.

I'm excited. One more feature will soon be available for my users; I have just enough time to refill my tea.

I look at Linear to check the status: work completed.

Cool.

I look at the output. No code. No PR. Nothing.

Instead, this message:

Feature deprioritized and cancelled. The proposed customer value in R-648 is not aligned with core problems identified during customer research.

My AI just told me No.

This system, specifically designed to always give what the user is asking for, refused to work on my feature.

So what just happened?

This wasn't a bug. This was by design. And it might be the future of how we build products. Here's the thing about most AI systems today: they're designed to be helpful. Almost pathologically so. Ask ChatGPT or Claude to build something, and they'll build it, even if they have to make up data to do it. I’m sure you're familiar with the term: hallucination.

But what if instead of one eager-to-please AI, you had a product team? One that could push back on your ideas, demand evidence, and sometimes tell you you're wrong?

That's what just rejected my feature.

How competing agents make better decisions

If I'd simply asked a coding agent to work on R-648, it would have quickly built everything, no questions asked. But I configured my Obi workspace differently. Before coding any new functionality, the platform runs a three-agent validation:

Agent 1: The Advocate. Argues in favor of building the functionality. It needs to build a business case anchored in data to support its arguments.
Agent 2: The Critic. Argues against building the functionality. Same requirement: it must ground its reasoning in customer research and rational analysis.
Agent 3: The Judge. Listens to both agents debate, weighs the evidence, and makes a final decision on whether to proceed.

The agents could have access to customer interviews, usage analytics, and the product roadmap. In our example, the critic agent found that the key problem we are solving is cashflow management, nothing related to long term investments. The advocate couldn't build a compelling case against this fact, so the judge decided to kill the feature.

I was annoyed for about thirty seconds. Then I realized: this is exactly what a good product team should do.

This is a relatively simple example of a multi-agent workflow, three agents in a single loop. But other systems are far more complex. Take the Trading Agents project, where twelve specialized agents collaborate to develop trading strategies, each bringing different analytical perspectives and constraints.

The opportunities here are endless. But today, implementing, testing, iterating, and scaling multi-agent systems is still quite complex, and only accessible to technical folks.

That's the problem I'm solving with Obi.

About Obi

Obi is one of my research projects. It’s currently available in a private release shared with a few trusted friends. My goal right now is to explore what's possible when you give users the ability to orchestrate teams of specialized agents.

At its core, Obi is orchestration infrastructure for cloud agents. Instead of managing individual AI assistants, you create multi-layer workflows (think orgs, not just employees). You might have a research team, a marketing team, and a QA team, all with different goals and constraints that keep each other in check.

Workflows are triggered by API calls or external tools integrated into the platform. For example:

Gmail trigger: When I receive an email from a customer, route it to my support squad for triage and response drafting
Google Calendar trigger: When a meeting starts, brief me on the participants and recent context
Plaid trigger: When a suspicious transaction hits my credit card, investigate and alert me
Stock price trigger: When a company I'm tracking hits a threshold, analyze whether it's time to act

These are simple examples. The real power comes when you chain these triggers together into sophisticated workflows where agents collaborate, debate, and make decisions.

For now, I'm just excited to see where this research leads. We're still early in understanding how to effectively orchestrate multiple AI agents.

But if my AI can save me from myself occasionally, I think we might be onto something....